Every university in Europe is currently running the same conversation, and it is the wrong one. The conversation is about permission: what may students use, what must they declare, how do we catch them. It produces policy documents, disclosure forms and detection subscriptions, and it produces almost no change in what is actually taught or how anyone actually learns. Meanwhile the evidence base — which has grown from nothing to several hundred studies in three years, and which this library assembles — has been quietly answering a different and far more consequential question: what does a machine that can explain anything, at any hour, at negligible cost, do to the production of an engineer?

The answer that emerges is not the one either camp expected. The optimists expected a tutor in every pocket and a two-sigma revolution; the pessimists expected a generation that cannot think. The measured reality is sharper and more useful than either. Where AI tutoring has been put through a proper randomised trial with a pedagogically constrained system — Kestin and colleagues’ Harvard physics experiment, published in Nature Scientific Reports, where students learned more than twice as much in less time than in an expert-taught active-learning class; the World Bank’s six-week Nigerian trial returning 0.31 standard deviations, among the most cost-effective learning interventions ever measured; Stanford’s Tutor CoPilot RCT across 900 tutors — the effects are large, real, and replicated across wildly different contexts. And where AI has simply been made available as a general-purpose assistant with no pedagogical constraint, the measured effect on thinking runs the other way: Microsoft Research and Carnegie Mellon’s study of 319 knowledge workers found that higher confidence in the AI predicts less critical-thinking effort, with users shifting from producing judgement to verifying someone else’s.

Those two findings are not in tension. They are the same finding stated twice. The variable that decides whether AI raises or lowers human capability is not the model — it is the structure wrapped around the model. Kestin’s tutor was explicitly forbidden to give answers; it was built to withhold, to question, to force retrieval. CS50’s duck at Harvard, the largest educational agent deployment on record at 211,000 students and 10 million queries, was engineered with the same intention — and its own published evaluation reports that 22% of responses contained code blocks despite instructions not to hand out solutions, rising to 48% at conversation level. That is the whole discipline in one number: the pedagogy is in the guardrail, and the guardrail leaks. An institution that buys models and skips the structure has bought the failure mode without the benefit.

This is why “should we allow it” is the wrong conversation for a technical university in particular. The prevalence question is settled: the HEPI/Kortext survey of UK students found 92% using generative AI, up from 66% a year earlier, and 88% using it in assessed work. MIT’s own institution-wide survey, reported by its Ad Hoc Committee in August 2026, found 46% of undergraduates using LLMs daily — and, more damningly, that while two-thirds of students believed AI would matter in their careers, only 25% felt MIT was adequately preparing them to use it. That gap is the actual institutional failure. It is not a discipline problem. It is a curriculum problem, an assessment problem, and above all a staff-capability problem, and every month spent litigating permission is a month not spent closing it.

There is a further reason a technical university cannot treat this as a general higher-education issue with a general higher-education answer. The disruption is not uniform across disciplines; it is concentrated almost exactly where a technical university lives. Computing education has the deepest evidence base and the sharpest dislocation — the ITiCSE working group’s landmark survey, Becker and colleagues’ “Programming Is Hard — Or At Least It Used To Be”, and the GitHub/Microsoft/MIT randomised trial in which Copilot users completed a programming task 56% faster — because code is the modality large language models are best at. The graduate attribute that a Faculty of Information Technology has spent decades certifying is precisely the one whose market price is moving fastest. Meanwhile Stanford’s Digital Economy Lab, tracking millions of payroll records, finds employment declining specifically for young workers in the occupations most exposed to AI — the entry-level software and technical roles that are the destination of a technical university’s undergraduates. The institution’s product is being repriced from both ends simultaneously.

And yet the same discipline structure that creates the exposure also supplies the defence, which is the most under-appreciated point in this entire literature. Engineering education already possesses, as its native pedagogical form, the thing every other faculty is now scrambling to invent: assessment against a physical or functional artefact that must actually work. A bridge calculation is checked by statics, not by an examiner’s impression of the prose. A circuit either oscillates or it does not. A robot either completes the task or falls over. CDIO, the international engineering-education framework to which the project-based assessment literature in this library belongs, has been building this for twenty-five years; the CESAER white paper on the engineer of the future, written by the very association of European technical universities that a school like ČVUT belongs to, argues for exactly this competence-and-challenge-based direction. Where the humanities must now reconstruct authenticity from first principles, engineering mostly has to stop drifting away from it — away from the worksheet, the boilerplate lab report, the individually-submitted problem set that a model completes in nine seconds.

So the reframe that organises this report is this. AI in teaching is not a tool question, it is a capability question — and the capability being built or destroyed is the institution’s, not the student’s. Universities that treat generative AI as software to be procured and policed will get the measured downside: offloaded thinking, unenforceable rules, an integrity arms race they lose, and graduates who use models badly because nobody taught them to use models well. Universities that treat it as a redesign of the teaching production function — what is assessed, what is taught, what staff can do, what data the institution keeps and what it is allowed to keep — get the measured upside, which is genuinely large, and get it disproportionately for their weakest students, which is where a technical university’s real losses are concentrated. The evidence for that last claim is the most decision-relevant thing in this library, and it is finding number two.

The findings in brief

• The tutoring effect is real, large and replicated — but only for systems built to withhold. Harvard’s constrained physics tutor beat expert-led active learning by more than 2×; Nigeria returned 0.31 SD; the pooled meta-analytic effect of ChatGPT on learning is g = 0.670. Unconstrained chatbots do not reproduce it.

• The gains concentrate in the weakest performers. Novices gained +34% in the NBER field study versus almost nothing for experts; Tutor CoPilot helped students of lower-rated tutors most. For an institution with 30%+ first-year failure, this is the single highest-value finding in the library.

• AI-text detection does not work, and building policy on it is an equity liability. Fourteen detectors failed systematic testing; GPT detectors flagged ~61% of non-native-speaker essays as AI-written. Assessment redesign is the only durable lane.

• Computing and programming education is the most disrupted subject on earth and has the best evidence about it — and the disruption is to learning objectives, not merely to cheating.

• Over-reliance is measurable, and it is the real cost — not plagiarism. Confidence in AI predicts reduced critical-thinking effort; students know it and are frightened of it (90% of MIT respondents concerned about overreliance).

• A deployed teaching agent is astonishingly cheap and demonstrably useful — and it leaks. CS50: $1.50 per student per year, 94% found it helpful, 22% of responses handed out code anyway.

• Faculty capability, not technology, is the binding constraint — and the frameworks to fix it (UNESCO, DigCompEdu, ETH Zurich’s lecturer framework) already exist and are unused.

• Engineering’s artefact-based pedagogy is a structural advantage nobody is exploiting — the lab, the design studio and the capstone are already AI-resistant assessment.

• The law has already decided most of this is high-risk. The EU AI Act’s Annex III names admission, outcome evaluation, level assignment and exam monitoring as high-risk uses — which is most of what a university would want to automate.

• Almost none of it has been evaluated to the standard the university would demand of its own research — and early-warning models specifically degrade across cohorts and subgroups.

How this report is organised

The ten findings are ranked by strength of evidence and decision-weight combined — how well the claim is established, and how much changes if you act on it. Four tests decide the ranking: whether the core result rests on randomised or quasi-experimental designs rather than surveys of opinion; whether independent teams in independent contexts converge on it; whether the outcome measured is learning or performance rather than satisfaction; and whether anything has been replicated at scale in a real institution rather than a laboratory. A result from a pre-registered RCT with a learning outcome outranks a beautifully-argued position paper. A deployment evaluation with published failure rates outranks a vendor case study. Sector surveys of what administrators intend rank last, however often they are quoted.

Each finding follows the same discipline: the claim in one line; the actual studies from the library that support it; an explicit statement of what the evidence does not show, because this field is drowning in over-claiming in both directions; and the institutional move it obliges. Where a finding bears specifically on a mid-sized European public technical university — the case that motivated this library — that is drawn out at the end of the section rather than allowed to colour the evidence. The companion report, The ČVUT Playbook, does the institution-specific work in full.

1. The tutoring effect is real and large — but only for systems built to withhold

Where AI tutoring has been tested properly, the effect sizes are among the largest in the history of education research. Every one of those trials used a system deliberately engineered not to answer.

Start with the strongest single study in the library. Kestin, Miller and colleagues ran a randomised crossover trial in Harvard’s PS2 physics course, published in Nature Scientific Reports in 2025. Students learned the same material either in a class taught by expert instructors using active-learning methods — already the gold standard against which everything else underperforms — or with an AI tutor built on GPT-4 and constrained by a detailed pedagogical prompt. Students in the AI condition learned more than twice as much, in less time, and reported higher engagement and motivation. The comparison is what makes it remarkable: this is not AI versus a bad lecture. It is AI versus the best-evidenced form of human undergraduate physics teaching, in one of the world’s most selective institutions, and the AI condition won on a pre-specified learning measure.

The result does not stand alone. The World Bank’s From Chalkboards to Chatbots working paper reports a six-week randomised trial of GPT-4-based virtual tutoring in Nigeria returning 0.31 standard deviations — which the authors, using standard learning-adjusted year conversions, place at the equivalent of one and a half to two years of business-as-usual schooling, and among the most cost-effective interventions ever measured in the education literature. Stanford’s Tutor CoPilot trial — the first RCT of a human–AI tutoring system in live tutoring, across 900 tutors and 1,800 students — found +4 percentage points of topic mastery overall at roughly $20 per tutor per year. An exploratory randomised trial in authentic UK classrooms, included here as the closest European deployment evidence, tested both learning effect and safeguarding simultaneously. And the meta-analytic anchor pools 35 experimental studies across 4,193 participants and reports g = 0.670 for ChatGPT’s effect on learning outcomes, with no significant publication bias detected.

For context on how good that is, the library carries both of the necessary baselines. Bloom’s 1984 paper — the source of the two-sigma claim that every AI-tutoring pitch deck implicitly invokes — established one-to-one human tutoring plus mastery learning as the ceiling nobody could afford. The pre-LLM reality check is the Ma, Adesope, Nesbit and Liu meta-analysis in the Journal of Educational Psychology: 107 effect sizes across 14,321 learners found intelligent tutoring systems outperforming large-group instruction and non-ITS computer instruction, but not outperforming individualised human tutoring or small-group instruction. Thirty years of intelligent tutoring systems bought a real but modest gain at high engineering cost per subject. The current generation is producing larger effects, in weeks, in domains nobody hand-authored.

Now the constraint that makes the whole finding conditional. In every trial above, the system was built to withhold. Kestin’s tutor was instructed to reveal one step at a time, to ask before telling, to refuse to produce the final answer. Tutor CoPilot does not tutor the student at all — it coaches the human tutor in real time. CS50’s duck at Harvard, whose own published evaluation is the most honest deployment document in this library, was likewise engineered to refuse solutions. There is no trial in this library in which handing students an unconstrained frontier model and letting them chat produced a learning gain of this magnitude. The pedagogy is not in the model. It is in the wrapper — the system prompt, the retrieval over the actual course material, the refusal policy, the scaffolding sequence.

What the evidence does not show. It does not show that AI tutors beat good teaching in general — Kestin’s design gave the AI condition a highly structured, expertly-prompted tutor in a single well-defined physics topic, which is a favourable case and was designed to be. It does not show durability: almost every study measures learning at or near the end of the intervention, and the library contains no strong evidence on retention months later. It does not show that gains survive when the tutor is built by a busy academic rather than a research team. And the meta-analytic g = 0.670 pools studies of highly variable quality with mostly short interventions and mostly proximal outcome measures — the standard conditions under which education effect sizes later shrink.

What it obliges. Stop procuring chatbots and start commissioning constrained tutors, subject by subject, grounded in the institution’s own course material, with the refusal behaviour specified as a pedagogical requirement rather than a safety afterthought. The unit of investment is not a licence. It is a course.

2. The gains concentrate in the weakest — and that is where a technical university’s losses are

Across every well-designed study in this library, the benefit of AI assistance is largest for the least expert and smallest for the most expert. For an institution losing a third of its first-year cohort, this is the most valuable finding here.

The pattern is extraordinarily consistent across domains that share nothing else. Brynjolfsson, Li and Raymond’s NBER field study of 5,179 customer-support agents found an average productivity gain of +14% issues resolved per hour — but +34% for novice and low-skilled workers, and minimal impact on experienced high-skilled workers. Their interpretation, which the data supports, is that the model diffuses the tacit knowledge of the best performers to everyone else. Stanford’s Tutor CoPilot trial found the same shape inside education: +4pp overall, but +9pp for students working with lower-rated tutors — the AI closed part of the gap between a weak tutor and a strong one. The Dell’Acqua, Mollick and Lakhani field experiment at BCG, run on 758 consultants, found the largest gains among below-average performers, who improved dramatically toward the group mean.

This is a compression effect, and it points somewhere very specific. A technical university’s most expensive, most persistent and least discussed failure is not the mediocrity of its best students. It is attrition in the first two years, concentrated in the gateway subjects — mathematical analysis, physics, mechanics, the first programming course — where a student who falls two weeks behind cannot recover because the only remediation available is a queue outside an office hour that clashes with another lecture. The published Czech numbers make the scale concrete: at ČVUT, first-year bachelor study failure ran at 31.8% university-wide in 2024, ranging from 9.5% at the Faculty of Architecture to 51.1% at the Faculty of Mechanical Engineering and 48.1% at Transportation Sciences. Half of one faculty’s incoming bachelor cohort fails in year one. Those are not students who lack capacity; they are overwhelmingly students who lacked a patient explanation at eleven at night in week six.

That is precisely the good that a constrained tutor supplies at negligible marginal cost, and it is precisely the population the evidence says benefits most. It is worth being blunt about the arithmetic, because it reframes the entire investment case. A technical university spends heavily on admissions marketing to fill a cohort, then loses a third of it to a failure mode that a well-built tutoring layer measurably addresses. The retention gain is worth more, in both money and mission, than every productivity saving in the administrative use cases that dominate university AI strategies.

Two cautions before anyone builds the business case. First, compression is not only good news: if AI lifts the floor without lifting the ceiling, the signal value of a degree compresses too, and the strongest students gain least from the institution’s investment. Second, the same compression logic that helps a struggling first-year is what makes over-reliance dangerous later — a student permanently held at the level the tool provides has been given a floor and a ceiling in the same object. Finding five deals with that directly.

What the evidence does not show. None of the compression studies are from engineering education specifically, and the two largest are from work settings rather than degree programmes. They measure task performance, not the acquisition of durable expertise — and the mechanism, diffusion of expert tacit knowledge to novices, is precisely the mechanism that could substitute for learning rather than produce it. No study in this library demonstrates that AI tutoring reduces university dropout. That trial has not been run, which is itself a finding, and it is exactly the trial a technical university with a 31.8% first-year failure rate is best placed in Europe to run.

What it obliges. Point the first serious deployment at the gateway courses with the worst failure rates, not at the flagship master’s programmes where the enthusiasts are. And instrument it as a trial, so the institution ends up owning evidence rather than anecdote.

3. Detection does not work — assessment redesign is the only durable lane

AI-text detection has been systematically tested and it fails; worse, it fails asymmetrically against non-native speakers. Any integrity policy resting on detection is both ineffective and an equity liability.

Weber-Wulff and colleagues, working through the European Network for Academic Integrity, tested fourteen AI-text detection tools under controlled conditions. The finding is unambiguous: the tools are neither accurate nor reliable, they skew toward classifying output as human-written, and their performance collapses under light obfuscation — machine translation, minor paraphrase, a pass through another model. This is not a maturity problem that a better product cycle fixes; it is close to information-theoretically inevitable as models converge on fluent, unremarkable prose.

The equity finding is worse and should end the argument on its own. Liang, Yuksekgonul, Mao, Wu and Zou at Stanford ran GPT detectors over TOEFL essays written by non-native English speakers and over essays by native-speaking US eighth-graders. The detectors were near-perfect on the native-speaker writing and misclassified roughly 61% of the non-native-speaker essays as AI-generated, with over half flagged by all seven detectors tested. The mechanism is that detectors key on lexical richness and syntactic variety — exactly the dimensions on which a competent second-language writer differs from a native one. For any technical university with a substantial international cohort, deploying such a tool means systematically accusing international students of misconduct at several times the rate of domestic ones, on the basis of their second-language fluency.

So the enforcement lane is closed. The library’s answer to what replaces it is unusually well-developed, because Australia’s regulator did the work first. TEQSA’s 2023 discussion paper Assessment Reform for the Age of Artificial Intelligence set out the two principles that now underpin most serious sector guidance worldwide, and its 2025 follow-up, Enacting Assessment Reform in a Time of AI, reports what institutions actually did with them. The core move is to stop treating every assessment as if it served one purpose. Some assessment exists to certify that a named human has a capability — and that requires secured conditions and identity assurance, at programme level rather than in every task. Everything else exists to develop capability — and there AI use should be open, taught and part of the point. Trying to make every assignment do both jobs is what produced the unwinnable arms race. QAA’s guidance adds the governance procedure: a four-step triage of which assessments actually need redesign, run through existing internal quality assurance rather than as a parallel emergency process.

For engineering, the practical translation is more favourable than for most disciplines, and the library supplies the specifics. The CDIO paper on project-based assessment in the era of generative AI reworks the PBL evaluation grid with explicit AI-use criteria — you grade the process, the design decisions and the defence, not only the artefact. The Integrevise research report on oral assessment gives the operating model, cost and staffing constraints of running vivas at cohort scale, which is the obvious authentication lane for a technical university and the one most often dismissed as impossible without checking the numbers. And ČVUT’s own Methodological Instruction 5/2023 already contains an activity-by-activity permitted / partly-permitted / forbidden schema — a more concrete instrument than most European universities possess, though written before the assessment-redesign literature matured and now due a revision that moves it from a rules document to a design document.

What the evidence does not show. It does not show that detection tools are useless in every configuration — they retain some signal on unedited long-form output, and Turnitin-class vendors dispute the specific error rates. It does not show that secured in-person assessment is unproblematic: TEQSA is explicit that identity assurance at programme level is hard, expensive and easy to implement badly. And there is no strong evidence yet on whether two-lane assessment actually preserves standards, because it is too new to have graduated a cohort.

What it obliges. Retire detection as a basis for misconduct proceedings, immediately and explicitly, and say why in public so that staff stop relying on it informally. Then run the triage: for every programme, identify the small number of points where certification genuinely requires secured conditions, secure those properly, and free everything else to be taught with AI in the open.

4. Computing education is the most disrupted subject on earth — and the disruption is to objectives, not to cheating

The subject a technical university teaches most confidently is the one large language models are best at. The published response from the computing-education research community is not about misconduct; it is about which learning objectives are still worth certifying.

No other discipline has responded to generative AI with this much empirical work this fast, which makes computing education the closest thing the sector has to a natural experiment. The ITiCSE working group report — twenty-odd authors, the landmark community survey in this library — covers code generation, code explanation, autograders, automated feedback, integrity and curriculum response in a single document, and its conclusion is that the pedagogical questions dwarf the disciplinary ones. Becker, Denny, Finnie-Ansley, Luxton-Reilly, Prather and Santos put the point in the title of their SIGCSE paper: Programming Is Hard — Or At Least It Used To Be. Their argument is that a large fraction of CS1’s traditional learning objectives were proxies. We never actually wanted students to be able to write a for-loop from memory; we wanted them to be able to decompose a problem, and writing the loop was how we checked. The proxy has broken. The underlying objective has not.

The industry-side evidence explains why the objectives must move rather than be defended. Peng, Kalliamvakou, Cihon and Demirer’s randomised controlled trial — GitHub, Microsoft and MIT — found developers using Copilot completed a standard programming task about 56% faster than the control group. Whatever a university thinks about AI in coursework, its graduates enter a profession where this is the baseline expectation. Certifying an ability to produce code unaided, slowly, is certifying a skill the employer will not buy.

The most useful study in this angle is the most uncomfortable. Prather and colleagues observed CS1 students actually using Copilot on a real assignment — the paper is titled, from a student quote, “It’s Weird That it Knows What I Want”. What they document is a set of genuinely new interaction pathologies: drift, where the student’s mental model of the problem quietly diverges from the code accumulating on screen; over-trust, where plausible output is accepted without verification; and a collapse of the metacognitive loop that novice programming is supposed to build. Ma, Chen and Konomi’s study of dialogue logs from a beginner Python course adds the typology, clustering student–ChatGPT interaction into four distinct usage patterns and linking each to performance — which means usage pattern, not usage volume, is the variable that matters and the thing worth teaching.

The instructional-response literature is thinner but concrete. The ASEE study on ChatGPT in programming courses documents the practice of requiring students to submit their own code alongside the AI’s and account for the difference — an assessment pattern that converts the tool into the object of study. The JITE paper supplies the instructor-side view of benefits and adverse impacts, which is what faculty development has to start from.

What the evidence does not show. There is still no strong longitudinal evidence on what happens to programming expertise across a whole degree under heavy AI use — the studies are single-course, single-semester, and mostly measure task outcomes rather than developed capability. The Copilot productivity RCT measured a well-specified task, not the messy comprehension-and-maintenance work that dominates real engineering. And the four-pattern typology is from one course at one university.

What it obliges. Rewrite CS1 and CS2 learning outcomes explicitly around decomposition, specification, verification, debugging and reading unfamiliar code — the objectives that survive — and assess those directly rather than through the broken proxy. Then treat every other engineering discipline as being roughly two years behind computing on the same curve, and start the same work now rather than waiting for its own crisis.

5. Over-reliance is measurable, and it is the real cost — not plagiarism

The strongest argument against casual AI adoption is not integrity. It is that confident use of a capable model measurably reduces the critical-thinking effort of the person using it — and students are more worried about this than their teachers are.

The central study is Lee and colleagues at Microsoft Research and Carnegie Mellon, published at CHI 2025: 319 knowledge workers supplied 936 first-hand examples of generative AI use at work. Two findings matter. Higher confidence in the AI predicts less critical-thinking effort; higher self-confidence in one’s own expertise predicts more. And the nature of the effort shifts — from information gathering and problem-solving toward information verification and response integration. That is not automatically bad: verification is real cognitive work and, done well, is exactly the skill finding ten of this report argues should be taught. It is bad when it is not done, and the study’s confidence finding says that the better the tool gets, the less likely the user is to do it.

This connects to a body of work that ENSI’s Education for the Agentic Age library treats at length under cognitive debt and cognitive offloading — the well-established finding that capability which is habitually externalised is not merely unused but progressively unavailable. In a professional formation context, that is the whole risk. An engineer who cannot check the model is not an engineer who is slower; they are an engineer who cannot tell when the answer is wrong, in a profession where being unable to tell is how people get hurt.

The most striking evidence that this is not an academic worry comes from students themselves. MIT’s Ad Hoc Committee report of August 2026 cites a survey of 1,002 affiliates in which 90% were somewhat or very concerned about overreliance, including 67% “very concerned”, and 45% very concerned about inaccurate outputs — this among a population where 46% of undergraduates use LLMs daily. Students are simultaneously the heaviest users and the most alarmed constituency. The institution-wide MIT survey found only 23% of respondents optimistic about generative AI, and a split verdict on self-efficacy: 40% felt AI made them more capable against 27% who felt it made them more replaceable.

The counterweight in this library is the complementarity literature, which is more sober than the enthusiasm it is usually cited to support. Hemmer and colleagues’ review formalises when a human–AI team actually beats either party alone and then reports the empirical record, which is frequently disappointing — genuine complementarity is harder to achieve than the framing suggests, and many studies find the team underperforming the better of its two members. Dell’Acqua and colleagues’ jagged-frontier experiment gives the sharpest single image: inside the frontier of tasks AI handles well, consultants produced work rated 40% higher in quality; on a task just outside it, they were 19 percentage points less likely to reach the correct answer than colleagues working without AI. The tool did not merely fail to help outside its competence. It actively degraded performance, because people could not see where the edge was.

What the evidence does not show. The Microsoft/CMU study is correlational and self-reported; it establishes an association between confidence and reduced effort, not that AI use causes cognitive decline. There is no longitudinal study in this library tracking engineering students’ capability across a degree under sustained AI use — the single most important missing evidence in the entire field. And the jagged-frontier result comes from consulting tasks, not technical ones.

What it obliges. Teach the frontier explicitly — where these systems are strong, where they fail, and how to tell which side of the line a given task sits on — as a required, assessed competence rather than an induction slide. And design at least one point in every programme where the student must demonstrate the underlying capability without assistance, not to catch anyone, but because a professional formation that never checks is not a formation.

6. The deployed teaching agent is cheap and useful — and it leaks

Two universities have run educational AI agents at real scale for years and published honest evaluations including their failure rates. The cost is trivially low, the reception is excellent, and the guardrails do not hold. All three facts must be planned for together.

Harvard’s CS50 is the largest deployment on record and the most valuable document in this library for anyone about to build something. By mid-November 2024 the CS50 Duck had been used by approximately 211,000 students, processing 10 million queries, at an average cost of $1.50 per student per year. The rollout was staged sensibly — 70 students in summer 2023, several hundred on campus that autumn plus thousands online — and the reception was strong: 75% of students used the tools frequently, 94% found them helpful and effective; in the earlier cohort, 53% “loved” and 33% “liked” them. An accuracy audit in that earlier paper found 22 of 25 curricular answers correct (88%) and 30 of 39 administrative answers correct (77%).

Then the failure data, which CS50’s team deserves considerable credit for publishing. Of the 10 million messages, roughly 2.1 million responses — 22% of all interactions — contained code blocks despite the system being instructed not to hand out solutions; at conversation level the figure is 48%, meaning about 635,000 of 1.3 million conversations involved the agent generating code. The team names the mechanism: instruction dilution, where a long conversation progressively erodes the system prompt’s authority until the pedagogical constraint stops binding. Their response was not a better prompt but a human-feedback loop with teaching fellows reviewing and correcting agent behaviour — which is to say the working architecture is not an agent, it is an agent plus a staffed quality process.

Georgia Tech’s Jill Watson provides the longer time series and the operational numbers. Across the OMSCS programme — an online master’s of roughly 9,000 students that at one point supplied about 9% of all US computer-science master’s degrees — Jill Watson Q&A served more than 4,000 students across more than a dozen classes and saved teachers more than 500 hours. Quality improved steeply with iteration: coverage rose from about 21% of questions at 80% precision in 2017 to over 96% coverage at over 86% precision by autumn 2019. Agent Smith, the companion system, generates a new Jill Watson for a fresh syllabus in about 25 hours of preparation. The LLM-era rebuild reports 76.7% answer accuracy against 31.3% for a generic OpenAI Assistants baseline on the same evaluation, with a 6.8-second average response time — the clearest available demonstration that grounding in course material, not model choice, is what makes an educational agent work.

What the evidence does not show. Neither deployment is a controlled trial: nobody randomised students into having the duck or not, so the learning effect is unmeasured and the satisfaction numbers cannot substitute for it. Both are computer-science courses at elite institutions with unusual engineering capacity. And the cost figures are inference costs — they exclude the staff time that both papers show is the actual requirement.

What it obliges. Budget for the humans. The published evidence says a course agent costs a couple of euros per student per year in compute and a permanent fraction of a teaching-assistant post in supervision — and that the supervision is what separates Georgia Tech’s 96% coverage from a broken pilot. Institutions that fund the licence and not the loop are buying the leak.

7. Faculty capability is the binding constraint — and the frameworks already exist

Every serious study of why AI adoption stalls in universities returns the same answer: not technology, not policy, not student resistance — staff capacity, incentives and time. The instruments to fix it are published, free, and largely unused.

Ithaka S+R’s interview study across nineteen North American universities is the most textured account of what actually happens: instructors adopt in isolated pockets, driven by individual enthusiasm, hitting institutional frictions that have nothing to do with the technology — no time to redesign a course, no recognition for doing so, no clarity on what is permitted, and no one to ask. The arXiv multi-institution study of barriers to generative AI adoption maps the same phenomenon across disciplines and roles, and its value is in showing the barriers are multi-level: individual confidence, departmental norms and institutional policy each block independently, so fixing one changes little. The EDUCAUSE landscape data shows institutions know this — training for faculty (63%) and staff (56%) top the list of AI strategic-planning elements — while the same sector’s Horizon Report describes faculty roles and workload as the variable most likely to determine outcomes.

What makes this finding actionable rather than merely gloomy is that the instruments exist. UNESCO’s AI Competency Framework for Teachers specifies fifteen competencies across five dimensions — human-centred mindset, ethics of AI, AI foundations and applications, AI pedagogy, and AI for professional development — at three progression levels. The European Framework for the Digital Competence of Educators (DigCompEdu), from the Commission’s Joint Research Centre, gives twenty-two competences in six areas on a six-level proficiency ladder, and is the instrument a European public university’s staff development is already legible against. And ETH Zurich’s AI Competence Framework for Lecturers is the one built by a peer technical university for its own academics — nine pages, organised by competence area and proficiency level, and effectively a ready-made curriculum that any European technical university could adapt in a term rather than draft in a year.

The gap between the availability of these frameworks and their use is the finding. Sector surveys report institutions planning training; the interview evidence reports academics who have never been offered any. A framework that is cited in a strategy document and never converted into scheduled, workload-credited, discipline-specific development produces nothing.

What the evidence does not show. There is no evaluation in this library of a faculty AI-development programme against a learning or teaching-quality outcome — the frameworks are consensus instruments, not tested interventions. The ERIC survey evidence on faculty technology use predates generative AI. And the barriers study is a preprint from a single multi-institution sample.

What it obliges. Treat academic AI capability as an operational deliverable with a named owner, a budget and a completion figure — not as a set of optional workshops. Adopt ETH Zurich’s lecturer framework rather than writing one. And make the entry point discipline-specific: a mechanical engineer will not attend a generic session on prompt writing, and should not be asked to.

8. Engineering’s artefact-based pedagogy is a structural advantage nobody is exploiting

The assessment form the rest of higher education is now scrambling to invent — work judged against something that must actually function — is engineering’s native mode. The sector’s problem is that it has been drifting away from it for thirty years.

The assessment-reform literature converges on a single prescription: make the object of assessment something a model cannot supply, which in practice means process, defence, iteration and a working result. Engineering already has all four in its lab, its design studio and its capstone. A structures calculation is checked against statics. A control loop either stabilises or oscillates. A robot completes the task or falls over. The verification is external to the assessor’s impression of the text — which is precisely the property that essay-based disciplines have lost and cannot easily rebuild.

The library’s engineering-specific material shows what exploiting this looks like. The CDIO conference paper on project-based assessment in the era of generative AI reworks the PBL evaluation grid with explicit AI-use criteria, giving a concrete rubric pattern for grading team projects where AI is permitted — the assessment is of design decisions and their justification, not of authorship. The ASEE study of image-generative AI inserted into conceptual design in a CAD class is the closest thing to a design-studio protocol for AI-assisted ideation. Martín-Núñez and Díaz Lantada’s review in the International Journal of Engineering Education supplies the taxonomy that keeps the strategy coherent: AI as subject matter to be taught versus AI as teaching infrastructure to be built — two different programmes, routinely conflated, requiring different owners and different budgets. The gAI-PT4I4 paper on generative AI plus low-fidelity digital twins with VR and retrieval-augmented generation is the strongest available template for scaling laboratory and simulation work when physical lab capacity is the constraint.

The peer-institution consensus documents matter here because they establish that this is the declared direction of European technical universities rather than one institution’s bet. CESAER’s Engineer of the Future white paper, written by the association of European universities of science and technology, argues for competence-based and challenge-based learning, lifelong learning and digital transformation as the shape of engineering education. EuroTeQ’s Framework of Qualifications — the alliance deliverable defining what a European engineering graduate must be able to do — is the concrete competence architecture into which AI competences can be inserted without inventing a parallel framework. Both are network commitments a member institution has already made and can simply execute against.

The uncomfortable half of this finding is that the advantage is being squandered. Artefact-based assessment is expensive in staff time, and three decades of expanding cohorts with flat teaching budgets have pushed engineering programmes steadily toward the cheap end — the individually-submitted problem set, the templated lab report, the multiple-choice test. Every one of those is now worthless as evidence of capability. The strategic point is that the response to generative AI and the response to the long erosion of engineering pedagogy are the same response, which makes the current moment an unusually good one to argue for the resources.

What the evidence does not show. There is no controlled evidence that project-based assessment resists AI better than written assessment — it is an argument from the nature of the task, not a measured result, and a team project can be substantially AI-generated in ways a busy assessor will not detect. The CDIO and ASEE studies are small, single-course and self-reported. And the digital-twin work is a proposed framework with limited deployment evidence.

What it obliges. Audit where each programme’s assessment actually sits on the spectrum from artefact-and-defence to submitted text, and move the balance deliberately — accepting that this costs contact hours and saying so in the budget rather than pretending redesign is free.

9. The law has already decided most of this is high-risk

The EU AI Act classifies as high-risk exactly the university uses an efficiency-minded administration would automate first. This is not a future compliance question; the obligations are in force and the sequencing consequence is immediate.

Annex III of Regulation (EU) 2024/1689 names education and vocational training explicitly. The high-risk categories cover AI systems used to determine access or admission to educational institutions, to evaluate learning outcomes, to assess the appropriate level of education a person will receive, and to monitor and detect prohibited behaviour during tests. Read that list against a typical university AI wish-list — automated admissions triage, automated grading, adaptive placement, remote proctoring — and the overlap is nearly total. High-risk classification brings risk management, data governance, technical documentation, logging, transparency, human oversight and accuracy/robustness obligations, and a provider-versus-deployer distinction that changes materially if the university builds rather than buys.

This is why the sequencing in the earlier findings is not merely pedagogically sound but legally forced. Tutoring, explanation, feedback-for-learning and staff support are not on the Annex III list. Grading, admission and proctoring are. An institution that starts with the low-risk teaching layer builds capability, evidence and trust while its compliance work matures; an institution that starts with automated grading because it looks like the biggest efficiency win has taken on the heaviest obligations first, with no institutional competence yet built.

Three further instruments define the operating envelope, and they are cumulative rather than alternative. The ESG — the Standards and Guidelines for Quality Assurance in the European Higher Education Area — govern how any change to teaching, assessment or grading must pass through internal quality assurance and survive external accreditation; an AI-mediated assessment change that has not been through programme-level QA is not merely risky, it is unaccredited. Data protection sits on top: the EDPS orientations on generative AI and the EUDPR, and the GDPR regime generally, apply in full to student data flowing through a commercial model. Jisc’s Code of Practice for Learning Analytics is the practical instrument that converts those obligations into institutional procedure — responsibility, transparency, consent, minimising adverse impact and stewardship — and remains the best available starting point for a university’s data-governance model.

The Czech and European policy frame supplies the funding logic rather than further constraint. The Czech National AI Strategy to 2030 carries education, skills and research pillars and is the national mandate a public university’s programme can be attached to; the Government Council’s 2024 analysis documents that the country has no dedicated digital-education strategy, which is both a gap and an opening for an institution willing to define the practice. The Commission’s Digital Education Action Plan and the joint OECD/EC AI literacy framework provide the vocabulary against which national and EU funding will judge proposals.

What the evidence does not show. The AI Act’s application to specific university configurations is genuinely unsettled — whether a university fine-tuning a commercial model becomes a provider, how the research exemption interacts with teaching deployments, and where the boundary sits between a feedback tool and an outcome-evaluation system are all live questions that guidance has not fully resolved.

What it obliges. Sequence deliberately by risk class, not by perceived efficiency. Put the AI Act’s high-risk register, the ESG and the data-protection regime into one governance instrument owned by one office, before the first pilot rather than after the third.

10. Almost none of this has been evaluated to the standard the university demands of its own research

A technical university applies rigorous evidence standards to everything except its own teaching. The AI literature is where that inconsistency becomes expensive.

The standards exist and are in this library. The What Works Clearinghouse Procedures and Standards Handbook sets out how the US government decides whether an education study counts as evidence — design requirements, attrition thresholds, baseline equivalence, effect-size computation. The EEF’s evaluator guide is the operational playbook: protocols, pre-registration, statistical analysis plans, implementation and process evaluation. Measured against either, the great majority of published claims about AI in higher education — including a large share of what circulates as best practice — would not qualify as evidence at all. They are satisfaction surveys, single-course reflections, and vendor case studies with no comparison condition.

The specific technical caution matters most for the intervention universities most want to build. Gardner and colleagues’ study of temporal and between-group variability in college dropout prediction shows early-warning model performance degrading across cohorts and across student subgroups. A model trained on last year’s students underperforms on this year’s, and underperforms unevenly — worse for the subgroups an equity-minded institution most wants it to serve. Since at-risk prediction is both the most attractive analytics application and an Annex III-adjacent use, this is the finding that should govern how it is deployed: as a trigger for offering help, monitored for subgroup drift, never as an input to a decision about the student.

The rest of the field’s methodological weaknesses are ordinary and predictable. The meta-analytic g = 0.670 pools mostly short interventions with proximal outcomes, exactly the conditions under which effect sizes shrink on replication. Deployment papers report satisfaction and usage, not learning. Almost nothing measures retention beyond the end of the intervention. And the studies with the best designs — Harvard’s physics trial, the Nigeria RCT, Tutor CoPilot — are, tellingly, the ones reporting the most disciplined interventions, which raises the possibility that design quality and intervention quality are correlated and the field’s average effect is inflated by neither being present.

The opportunity in this is larger than the caution. The gap between what is claimed and what is established is wide, the population needed to close it sits in every gateway lecture theatre, and a technical university already employs the statisticians. An institution that instruments its own deployment as a trial ends up owning evidence that nobody else has — publishable, fundable, and a durable reputational asset in a field where almost everyone is guessing.

What it obliges. Treat every AI teaching deployment as a study: comparison condition, pre-registered outcome, learning measure rather than satisfaction measure, subgroup analysis by default. It costs little more than doing it badly and produces something the institution can defend.

What to do first

The ten findings collapse into a short sequence, and its order is not arbitrary — it runs from highest evidence and lowest regulatory risk to the reverse.

First, close the detection question and open the assessment question. Withdraw detection from misconduct procedure explicitly and publicly, then run the TEQSA/QAA triage across every programme to identify the minimum set of points where certification genuinely requires secured conditions. This costs nothing, ends an unwinnable and inequitable enforcement effort, and is the precondition for everything else.

Second, build one constrained tutor for one gateway course with a bad failure rate — grounded in the actual course material, engineered to withhold, supervised by a named teaching assistant, instrumented as a randomised trial with a learning outcome. The evidence says the effect will be largest exactly there. Budget two euros per student per year for inference and a fraction of a post for the loop, because the published deployments say the loop is what works.

Third, fund academic capability as an operational deliverable. Adopt ETH Zurich’s lecturer competence framework rather than drafting one, map it onto DigCompEdu for European legibility, deliver it discipline-by-discipline with workload credit, and publish the completion figure. Every study of stalled adoption points here.

Fourth, move the assessment centre of gravity back toward the artefact and its defence — the direction CESAER and EuroTeQ have already committed European technical universities to, now with a second and more urgent justification.

Fifth, put the AI Act, the ESG and data protection into a single governance instrument before the first high-risk pilot, and sequence deployments by risk class rather than by perceived efficiency.

The reframe to hold, against a sector conversation that will keep pulling toward tools and rules: the capability at stake is the institution’s. The models are a commodity available to every student for the price of a coffee, and no policy will change that. What is not a commodity is a university that knows which of its assessments still mean anything, whose academics can teach with and against these systems, that catches its struggling first-years before they leave, and that can prove any of it. That is buildable, the evidence says roughly how, and — the one genuinely urgent point in this report — the institutions that start now will be the ones producing the evidence everyone else cites in three years.

The argument, before the list

The received story about artists is a story about temperament. On this account, some people are born with a strange sensibility — they see more, feel more, tolerate more chaos — and art is what that sensibility does when left alone. The story is convenient for everyone: it lets artists claim a mystery, lets scientists dismiss one, and lets educators file “creativity” under things that cannot be scheduled. The laboratory record, assembled here from 175 primary documents, says the story is wrong in a specific and useful way. Artistic thinking is not a temperament. It is a trainable cognitive stack — a set of distinct, measurable operations in perception, process and cognition that trained artists demonstrably run differently from everyone else, and that untrained people begin to run differently after training.

The evidence is not anecdote. It is drawing-accuracy experiments and eye-tracking traces (Cohen and Bennett’s classic isolation of misperception as the main source of drawing error; Cohen’s finding that gaze frequency predicts drawing accuracy), longitudinal neuroimaging of art students whose brains restructure across a course of training (Schlegel and colleagues’ “The Artist Emerges”), fMRI of jazz musicians and freestyle rappers generating new material inside a scanner (Limb and Braun; Liu and colleagues), corpus analyses of 150 Hollywood films across 70 years (Cutting, DeLong and Nothelfer), network-science mapping of semantic memory in high-original individuals (Kenett, Anaki and Faust), and a meta-analytic literature — on incubation (Sio and Ormerod), on personality (Feist), on the prefrontal geography of idea generation (Gonen-Yaacovi and colleagues) — that has quietly converged while the temperament myth held the stage. Where the artist’s advantage has been looked for in the eye itself, it has not been found: Perdreau and Cavanagh’s test of the old claim that artists “see their retinas” came back negative. The advantage lives higher up — in trained attention, trained selection, trained process. Which is precisely what makes it transferable.

Why should anyone outside the studio care? Because originality has become the binding constraint on the systems that matter. The bibliometric record of science itself — angle twelve of this library — shows papers and patents becoming measurably less disruptive over six decades (Park, Leahey and Funk), novel work systematically under-rewarded at first (Wang, Veugelers and Stephan’s bias-against-novelty analysis), while the work that ultimately lands hardest is the work that injects atypical combinations into conventional cores — Uzzi and colleagues’ finding across 17.9 million papers. Science, in other words, has a documented originality shortage and a documented originality premium, simultaneously. Any institution that could reliably increase its supply of original cognition would be buying the scarcest input in the knowledge economy at exactly the moment machine systems are commoditising the conventional kind.

And there is a population that has been training original cognition, deliberately and for centuries: artists. The transfer evidence makes the connection more than poetic. Root-Bernstein and colleagues’ studies of Nobel laureates found them many times likelier than average scientists to sustain serious arts and crafts avocations — a result extended in their 2019 PNAS-line study of STEMM professionals, where arts, crafts and design practices track scientific achievement. Ramón y Cajal, the founding draughtsman of neuroscience, is the canonical case — a recent essay in the library traces how his drawing practice shaped what he could see down a microscope. The correlational nature of this evidence has limits, and this report will be honest about them — the OECD’s Art for Art’s Sake review found the causal case for far transfer weak — but the direction of the signal is consistent: the people who change science disproportionately think like artists on the side.

So this report does something deliberately unromantic. It treats “how artists think” as an empirical question with an answerable structure, and ranks what the evidence actually supports. At the top sit findings with replicated, objective, convergent support — trained perception, the filmmaker’s command of attention, the oscillating architecture of the creative process. In the middle sit the cognitive mechanisms — far association, improvisation as a releasable brain state, problem finding, combinatorial imagination. At the bottom sit the findings whose effects are large but whose causality is weakest — personality, embodiment, and transfer itself. Nothing here requires belief in genius. Everything here is a difference someone measured.

The reframe to hold through all ten findings is this: what artists possess is not a gift but a stack — trained seeing, trained problem-formulation, trained oscillation between generating and judging, a loosened associative network, a practised ability to switch off self-censorship on demand, and a body and studio arranged as thinking instruments. Every layer of that stack has been measured. Several have been trained in non-artists inside controlled studies. That is the practical meaning of the evidence: originality is a capability, not a lottery — and the artist is its best-documented working model.

The findings in brief

• Artists see differently — and the seeing is trained, not given. Drawing accuracy is limited by perception, not the hand; artists’ eye movements, salience resistance and shape perception all differ measurably; training changes the brain (Cohen & Bennett · Chamberlain & Wagemans · Perdreau & Cavanagh · Schlegel).

• Filmmakers are running experiments on attention — and winning. Edit blindness, gaze synchrony, event segmentation and 70 years of corpus evolution show cinema is applied cognitive science (Smith · Hasson · Zacks · Cutting · Cohn).

• The creative process is an oscillation between generative and evaluative modes, visible in the brain and in artists’ own accounts (Ellamil · Sowden, Pringle & Gabora · Beaty · Botella).

• Original ideas come from far associations. High-original people carry measurably more flexible semantic networks, and originality rises with semantic distance and with time-on-task (Kenett · Benedek & Neubauer · Beaty & Silvia · Green).

• Improvisation shows originality is a releasable brain state — self-monitoring down, self-expression up — and improv training raises originality in ordinary teenagers (Limb & Braun · Liu · Hainselin).

• Artists find problems before they solve them. Problem construction measurably improves creative outcomes; studio process is discovery-shaped, not execution-shaped (Reiter-Palmon · Botella · Tversky · Scotney).

• Analogy, metaphor and blending are the combinatorial engines — the same machinery Dunbar filmed in world-class science labs (Gentner · Holyoak · Fauconnier & Turner · Dunbar · Uzzi).

• The original personality exists and is measurable — openness to experience is its strongest marker, split revealingly between artistic and scientific forms (Feist · Kaufman · Barron · Kyaga).

• The body is part of the thinking. Incubation works, walking boosts ideation, choreographers think with their limbs, and groups generate what no member holds (Sio & Ormerod · Oppezzo & Schwartz · Kirsh · Sawyer).

• Artistic originality transfers — with honest limits. Nobel-laureate avocations and STEMM correlations are strong signals; causal far-transfer evidence remains weak (Root-Bernstein · OECD · Schellenberg).

How this report is organised

The ten findings are ranked by strength and convergence of evidence, judged on four tests: whether the core result has been replicated; whether independent methods converge on it (behavioural experiment, eye-tracking, neuroimaging, corpus analysis, meta-analysis); whether the measures are objective rather than self-reported; and whether anything causal — a training study, an intervention — anchors it. A finding built on decades of replicated experiments with objective measures outranks one built on a single elegant fMRI study; a mechanism confirmed by meta-analysis outranks a beautiful theory; correlational biography, however striking, ranks last no matter how large its effect sizes. Each finding follows the same discipline: the claim in one line, the actual studies from the library, an explicit account of what the evidence does not show, and what a scientist, founder or policymaker should do with it. The library’s fifteen angle indexes are the bibliography; every study named below is in them.

1. Artists see differently — perception is trained, not given

The best-documented difference between artists and everyone else is perceptual: artists have retrained what they see, and the retraining is measurable at every level from eye movement to brain structure.

The foundation stone is Cohen and Bennett’s 1997 study, bluntly titled “Why can’t most people draw what they see?” By experimentally isolating the candidate failure points — misperception of the object, inability to make good representational decisions, misperception of one’s own drawing, motor incoordination — they showed that the main obstacle is the first one: people draw badly because they see conventionally. The hand is largely innocent; the eye is guilty. Chamberlain and Wagemans’ comprehensive review of the genesis of drawing errors traces the same conclusion across the perception, memory and motor stages of drawing, and the Kozbelt line of work — updated and replicated in Chamberlain, Drake, Kozbelt and colleagues’ “artists as experts in visual cognition” study — shows that trained artists outperform non-artists on visual tasks well beyond drawing itself. Robles, Bies, Lazarides and Sereno’s 2022 Scientific Reports study closes the loop in the modern era: veridical shape perception measurably tracks drawing ability.

The eye-tracking evidence shows how the trained eye works. Cohen’s 2005 study found that trained artists shift gaze between object and drawing far more frequently than novices — and that gaze frequency itself predicts drawing accuracy, as though expertise consisted partly of refusing to let memory adulterate the percept. Park, Williams and Chamberlain found artists’ global-to-local saccade ratios while drawing differ systematically from non-artists’. Koide and colleagues’ PLOS ONE study delivers the sharpest formulation: when experts view abstract paintings, their fixations are driven far less by bottom-up visual salience than novices’ — the trained eye sees top-down, going where its owner’s questions direct it rather than where the stimulus shouts. Francuz and colleagues found the same expertise signature in fixation patterns during compositional judgement.

Two further results give the finding its edge. First, Perdreau and Cavanagh’s direct test of the romantic hypothesis — that artists somehow access a raw, less-interpreted “retinal” image — came back negative: artists’ advantage is not low-level vision. It is learned control over attention, selection and constancy, which is exactly why it is trainable rather than congenital. Second, Schlegel and colleagues’ longitudinal NeuroImage study followed art students across their training and watched neural structure and function change as drawing skill grew — the artist, as their title has it, emerges. Perception here is not a window but an instrument, and instruments can be re-machined.

The neuroaesthetics canon shows what this trained seeing plugs into on the receiving side. Vessel, Starr and Rubin found that the most intensely moving aesthetic experiences engage the default-mode network — art reaching the systems of the self, not merely the eyes; Ishizu and Zeki located the experience of beauty, across visual and musical modalities, in medial orbitofrontal activity; and Chatterjee and Vartanian’s aesthetic triad — sensory-motor, emotion-valuation, meaning-knowledge — maps the full circuit an artwork traverses. Massaro and colleagues show viewer expertise reshaping even how paintings are visually explored. Perception in art is never just optics; it is optics wired to valuation and selfhood — which is why retraining it changes more than drawing.

What the evidence does not show: it does not show that artists have better eyesight, faster visual processing, or any general perceptual superiority — the advantage is specific to trained operations of attention and selection. Nor does the correlational strand (shape perception tracking drawing skill) by itself prove direction; the longitudinal and training evidence carries the causal weight.

The “so what” is the most direct in this report, because the tool is sitting on the desk. Fan, Bainbridge, Chamberlain and Wammes’ Nature Reviews Psychology review positions drawing as a versatile cognitive tool that changes what its user perceives, remembers and communicates — a claim with Cajal as its historical proof. A scientist or founder who learns observational drawing is not acquiring a hobby; they are running the best-validated perceptual retraining programme on record. If your work depends on noticing what the conventional eye smooths over — anomalies in data, unmet needs in a market, the detail everyone’s schema deletes — the artists’ evidence says: noticing is a skill with a syllabus.

2. Filmmakers run experiments on attention — film is applied cognitive science

A century of filmmaking craft encodes a working science of human attention — and when cognitive scientists finally tested it, the filmmakers were right.

This finding ranks second because its evidence is unusually convergent: eye-tracking, fMRI, behavioural change-detection and corpus analysis all land on the same conclusion from independent labs. The founding document is Hasson and colleagues’ 2008 neurocinematics paper, which put viewers of different films in a scanner and measured intersubject correlation — how similarly different people’s brains respond, moment by moment, to the same footage. The result: films differ dramatically and measurably in how tightly they control viewers’ brains. Directorial control is not a metaphor. It is a quantity, and some directors have more of it.

Tim Smith’s programme at Birkbeck explains the mechanism. His attentional theory of cinematic continuity formalises what editors discovered by feel: continuity editing works because it exploits the machinery of visual attention, placing cuts where the visual system is already committed to a saccade, a motion, an expectation. His edit-blindness eye-tracking study with Henderson showed the astonishing consequence — viewers simply fail to see cuts that follow the continuity rules, missing splices in the visual stream that a naive theory of perception says should be jarring. His later work with Martín-Portugués showed how match-action motion and audio timing manufacture global change blindness across cuts. Editors, in other words, spent a century converging on the parameters of change blindness decades before psychology named the phenomenon — a craft tradition that discovered real cognitive law by iteration, the way medieval builders discovered statics.

The corpus evidence shows the discovery process operating at industrial scale. Cutting, DeLong and Nothelfer analysed 150 Hollywood films across 70 years and found shot lengths evolving, decade by decade, toward the 1/f temporal rhythms characteristic of natural human attention — an entire industry performing a slow, unplanned gradient descent onto the attention dynamics of its audience. Cutting’s companion corpus work maps how filmmakers construct narrative space shot by shot. Zacks and colleagues’ fMRI work shows viewers’ brains segmenting narrative film into events precisely at the boundaries editors chose, and Magliano and Zacks show continuity editing shaping those segmentation networks during commercial film viewing. Cohn’s visual narrative grammar completes the picture from the linguistics side: sequential visual storytelling has a hierarchical constituent structure — a real grammar, extended in his later work from static sequences to film — which is why some cut orders parse and others do not; his tutorial paper turns the grammar into a practical toolkit for analysing any storyboard or sequence. Smith’s psychocinematics chapter closes the methodological loop, using eye-tracking to test film theory’s own claims about how directors steer the gaze — a hundred years of critical assertion suddenly exposed to falsification, and much of it holding.

What the evidence does not show: that filmmakers hold this knowledge explicitly — the craft knows things its practitioners cannot state, which is exactly what makes it interesting as a model of embodied expertise. And Loschky and colleagues’ “tyranny of film” study draws an important boundary: strong gaze synchrony does not guarantee shared understanding — where viewers look converges far more than what they comprehend. Attentional control is not semantic control.

The “so what” runs in two directions. For anyone who designs experiences that must pass through human attention — products, interfaces, briefings, lectures — cinema is a validated engineering literature, not an entertainment. Its principles (guide the eye before the cut; ride prediction rather than fighting it; segment events where the mind already segments) are testable and tested. For research strategy, the deeper lesson: mature craft traditions are unread datasets. Editing anticipated change-blindness research; drawing instruction anticipated perceptual-expertise research. A foresight-minded institution should ask which of today’s craft communities — game designers, prompt engineers, standup comedians — are currently sitting on the next such body of pre-scientific cognitive law.

3. The creative process is an oscillation between generating and evaluating

Creation is not one mental mode but a disciplined alternation between two — and the brain dissociates them cleanly enough to see in a scanner.

The anchor study is Ellamil, Dobson, Beeman and Christoff’s 2012 NeuroImage experiment, which put participants through cycles of designing book covers — generate, then evaluate, then generate again — inside an fMRI scanner. Generation preferentially engaged medial temporal regions associated with associative memory and novel combination; evaluation recruited executive regions together with default-mode regions, suggesting creative judgement is not cold quality control but a hybrid of analysis and imaginative simulation. The two modes are neurally dissociable, and the creative process consists of moving between them — not of occupying some single “creative state”.

The behavioural and theoretical record converges. Sowden, Pringle and Gabora’s dual-process account describes creation as shifting between associative and analytic modes, with skill lying in the timing of the shifts. Botella, Zenasni and Lubart’s study of art students’ own process accounts finds the same shape from the inside: real artistic process is iterative and looping — idea, test, judgement, re-entry — rather than the tidy linear pipeline of textbook stage models. Beaty and colleagues supply the network-level mechanism: their Scientific Reports fMRI study found that creative idea production is supported by coupling between the default network and the executive control network — two systems textbook neuroscience once treated as antagonists, working in alternating cooperation — a result consolidated in Beaty’s 2019 review of the creative brain’s network architecture. De Pisapia and colleagues found the same default-executive cooperation in professional artists planning a real artwork, which matters: the oscillation is not an artefact of laboratory tasks. It is visible even inside a single act of making — Miall, Nam and Tchalenko’s fMRI study of portrait drawing, from the team behind the celebrated eye-and-hand studies of the painter Humphrey Ocean, catches the drawing brain cycling between looking, deciding and executing rather than running one continuous programme.

Chrysikou and colleagues’ matched-filter hypothesis adds the report’s most counterintuitive mechanism: prefrontal cognitive control is not simply good — it is task-matched. High control serves well-defined problems; low control measurably benefits open-ended generation, which is why deliberate effort so often strangles the very ideas it is trying to produce. And Gonen-Yaacovi and colleagues’ ALE meta-analysis of 34 fMRI studies locates creativity’s generative and combinatorial operations across rostral and caudal prefrontal regions — solid meta-analytic ground under the claim that generation and evaluation are distinct computations, not moods.

What the evidence does not show: any single “creativity centre”, and emphatically not the folk right-hemisphere story. Boccia and colleagues’ meta-analysis finds musical, verbal and visuospatial creativity resting on partly domain-specific networks — creation is an orchestration, differently cast per domain.

The “so what” is organisational as much as personal. Most knowledge work runs generation and evaluation simultaneously — the meeting that brainstorms and critiques in the same breath, the writer who edits each sentence as it appears — which the matched-filter and dual-process evidence identifies as running both modes at once and doing each badly. Artists’ studio practice institutionalises the separation: sketch phases where nothing is judged, crit sessions where everything is. A lab or company can copy this directly — separate generative sessions from evaluative ones in time, place and even personnel, and treat the switch as the skill to train. The oscillation is the process; scheduling it is management.

4. Original ideas come from far associations — and the network is measurable

Highly original people carry differently structured semantic memories — more flexible, better connected across distant regions — and originality rises measurably with associative distance.

This is the mechanism-of-record for where new ideas come from, and its modern evidence is unusually elegant. Kenett, Anaki and Faust applied network science to the semantic memory of low- and high-creative individuals and found the high-creative networks measurably more flexible and interconnected — shorter paths between far-flung concepts, less rigid clustering. An original mind is not a bigger warehouse; it is a better-connected graph, in which “remote” ideas are simply less remote. This gives structural teeth to Mednick’s old associative theory — that creative individuals have flat rather than steep associative hierarchies, so unusual associates are nearly as available as obvious ones — which Benedek and Neubauer tested directly and refined: creative people’s advantage lies substantially in more effective search through associative memory, not merely a different gradient.

The process evidence shows far association operating in time. Beaty and Silvia’s serial-order work documents that ideas become measurably more creative the longer one generates — the first answers are everyone’s answers; originality lives past the point where most people stop. Beaty, Silvia, Nusbaum, Jauk and Benedek’s Memory & Cognition study shows divergent production riding on both associative and executive processes — the loose network plus the disciplined search of it — and Benedek’s work in Intelligence maps how executive functions serve both intelligence and creativity, against any romantic opposition between the two. On the neural side, Green and colleagues found that semantic distance in analogical reasoning parametrically engages frontopolar cortex — the brain treats far connection as its own operation. Kounios and Beeman’s insight programme — the Aha! moment and their Annual Review synthesis — shows solutions-by-insight have distinct neural signatures, with brain states preceding a problem predicting whether it will be solved by insight or analysis. Preparation shapes originality before the problem arrives. And Beaty and Johnson’s SemDis platform now scores associative originality automatically as semantic distance — the construct is measurable enough to be computed.

What the evidence does not show: that divergent-thinking scores are destiny. The library’s 2022 review of the Torrance Tests lays out the predictive-validity debates honestly — test-measured divergence correlates with, but does not guarantee, real creative achievement, which needs domain skill, motivation and opportunity (Amabile’s componential model, from the foundations angle, is the standing corrective).

The “so what”: originality has a substrate you can build. Feed the network heterogeneous material — distant fields, unshared experiences, deep non-work domains — because association can only connect what is stored. Then work the search: persist past the first wave of ideas (the serial-order effect is a free lunch almost nobody eats), and measure output by semantic distance, which SemDis-class tools now make practical for teams. A founder’s differentiated insight and a scientist’s atypical combination are, mechanically, the same event: a traversal between regions of the graph that competitors’ graphs do not connect. And because Kounios and Beeman show pre-problem brain states predicting insight, the preparation is not metaphorical — the mood, attention and expectation a person carries into a problem are part of the solving machinery, and can be set deliberately.

5. Improvisation shows originality is a releasable brain state

When trained improvisers generate in real time, the brain measurably reconfigures — self-monitoring down, self-expression up — and the state can be trained into ordinary people.

The seminal document is Limb and Braun’s 2008 PLOS ONE study, which put professional jazz musicians on a custom keyboard inside an fMRI scanner and compared improvisation against memorised performance. Improvisation came with a characteristic prefrontal reconfiguration — broadly, a retreat of the dorsolateral regions associated with deliberate self-monitoring and control, alongside engagement of medial prefrontal regions associated with self-expression. Liu and colleagues then found the same dissociated medial-versus-dorsolateral prefrontal pattern in an entirely different art form, freestyle rap, comparing improvised with rehearsed lyrics. Two genres, two labs, one signature: the generative state is partly a release — a temporary standing-down of the very supervision that ordinary cognition works so hard to maintain. This is the matched-filter hypothesis (finding 3) caught live: less top-down control precisely when the task is open-ended generation.

The surrounding evidence widens the base. Saggar and colleagues’ Pictionary-style fMRI study of improvised drawing found cerebellar-cortical dynamics supporting spontaneous figural creativity — recruiting the brain’s motor-automation machinery for idea generation, a hint that fluent generation behaves like a trained skill rather than deliberate reasoning. Norgaard and colleagues’ study of artist-level jazz improvisers maps the functional network connectivity of the practised improvising brain, and Arkin, Przysinda and Loui find structural grey-matter correlates of improvisational creativity — the state leaves a trace, as states that are trained do. Pressing’s foundational cognitive model of improvisation explains what the years of practice build: automated generative structures that free attention from execution — the release is affordable only because the underlying craft has been drilled to the point of costing nothing. And the crucial causal anchor comes from outside music: Hainselin and colleagues showed that an 11-week improvisational-theatre programme raised divergent-thinking originality and flexibility in ordinary teenagers — the state is not a property of elite musicians but a trainable disposition, teachable by curriculum. The applied-improvisation literature in the library explains what such training actually rehearses: the theatre’s “Yes, and” rule is deferral of judgement made into a bodily reflex — acceptance first, evaluation later — which is the generative half of finding 3’s oscillation, installed as habit.

What the evidence does not show: the neuroimaging samples are small and expert-heavy, as first-generation scanner studies are; the precise prefrontal geography varies across studies and tasks; and “deactivation” is a simplification of a reconfiguration that differs between musical and verbal improvisation. The honest claim is not a single switch in the head, but a replicated family of state changes with the same functional meaning: supervision loosens, generation flows.

The “so what”: most professional environments are engineered to keep the monitoring system permanently on — evaluation, status, audit. The improvisation evidence says original output needs sanctioned intervals where it is off, and that the off-switch strengthens with practice. Improv classes for scientists and founders are not team-building whimsy; they are state training with a controlled study behind them. The practical designs follow directly: regular low-stakes generative sessions whose output is explicitly unjudged; warm-up rituals borrowed from performers; and a personal practice — musical, verbal, physical — that rehearses the release itself.

6. Artists find problems before they solve them

The most distinctive move in the artistic process happens before solving begins: original creators treat the problem itself as the thing to be discovered, and problem construction measurably improves creative outcomes.

This is the oldest process finding in the library’s lineage — the problem-finding tradition that began with Getzels and Csikszentmihalyi’s studies of art students, carried forward in the library by its modern heirs. Reiter-Palmon and Murugavel’s review of problem construction consolidates the experimental record: when people actively construct and reformulate a problem before solving — rather than accepting it as given — their solutions are reliably more original and more effective, and problem-construction ability tracks creative performance across studies. Botella, Zenasni and Lubart’s study of art students’ own creative process confirms the shape from the practitioner’s side: the early phases of artistic work are dominated by searching, reframing and defining, not executing. The artist’s question is not “how do I solve this?” but “what is actually worth making?” — and the evidence says the quality of the eventual answer is substantially decided there.

The surrounding studies show how the finding operates in the wild. Tversky’s synthesis of the sketch-cognition programme documents designers and architects discovering problems in their own sketches — drawing something, then seeing in the drawing relations and possibilities they did not knowingly put there; the sketch is a problem-finding instrument, not a record of a finished thought. Glăveanu and Lahlou’s subjective-camera study — head-mounted cameras on working craftspeople — catches the same continuous renegotiation between maker and material in real studio practice. Scotney and colleagues add a structural result: creative inspiration disproportionately crosses domains in the early, idea-finding phase of the process — the problem-finding window is precisely when far material (finding 4) enters. And Dunbar’s in-vivo studies of molecular biology laboratories show the scientific analogue: the labs that generate discoveries treat anomalies not as noise to be explained away but as problems to be adopted — problem finding as institutional reflex.

Two further library strands widen the claim. Candy’s guide to practice-based research formalises what the studio tradition has long asserted: making is itself a method of inquiry, generating knowledge that could not have been specified in advance — the epistemic dignity of problem finding, written into research methodology. And Glăveanu’s sociocultural work — the Five A’s framework and his ethnography of a living craft community — shows that problem finding is rarely solitary: what counts as a problem worth working is negotiated continuously with materials, audiences and traditions, which is why isolated brainstorming so often produces problems nobody has.

What the evidence does not show: much of the tradition rests on interviews, self-report and small samples; the classic longitudinal claim — that problem-finding art students became the more successful artists years later — is suggestive rather than definitively replicated, and the modern experimental work measures near-term solution quality, not careers. Rank six reflects exactly this: a coherent, repeatedly supported process claim whose causal spine is thinner than the perceptual and neuroimaging findings above it.

The “so what” may still be the highest-leverage item in this report, because institutions systematically pay for the opposite. Grant systems, sprint plans and OKRs all reward solving pre-formulated problems on schedule; almost nothing rewards the reformulation step where originality is decided. The artist’s discipline transfers directly: budget explicit problem-finding time before committing to any solution path; require competing formulations of any important problem before work starts; treat a founder’s problem statement — not the product — as the primary creative artefact, iterated as many times. Where the reflex cannot be trusted, it can be scheduled. The science-funding evidence in the library shows this scales: Azoulay, Graff Zivin and Manso found that HHMI-style long-horizon funding — which tolerates early failure and does not demand a pre-specified deliverable — causally increases breakthrough scientific output relative to conventional grants. That is problem-finding time, purchased at institutional scale, with a measured return.

7. Analogy, metaphor and blending are the combinatorial engines

The machinery that makes new ideas out of old ones is not mysterious — it is analogy, metaphor and conceptual blending, three well-theorised operations that artists work harder than anyone else.

The theoretical spine here is among the strongest in cognitive science. Gentner’s structure-mapping theory formalised what an analogy actually is — an alignment of relational structure, not surface resemblance — and her later synthesis with Markman extends the account across analogy and similarity. Holyoak’s UCLA programme built the parallel theory of analogical and relational reasoning; Hofstadter’s essay presses the maximal claim that analogy-making is the core of cognition itself. Lakoff and Johnson’s conceptual-metaphor work shows the machinery is not an occasional ornament but the pervasive structure of everyday thought — abstract thinking runs on metaphors drawn from bodily experience. Fauconnier and Turner’s conceptual-blending theory describes the general operation: multiple input spaces projected into a blend with emergent structure that belongs to neither input — the closest thing on record to a formal mechanism of the genuinely new. Schacter, Addis and Buckner ground the engine in memory: imagination is constructive episodic simulation, recombining stored experience into scenarios never lived — the same system that remembers is the system that invents. That is Vygotsky’s near-century-old thesis vindicated by neuroscience: his classic essay on imagination in childhood already argued that all imagination is recombination of experienced elements, so the richer the experience, the further the reach. Byrne’s rational-imagination work shows even counterfactual thought is generated by systematic principles rather than free fancy — and the aphantasia research in the library (Dawes and colleagues’ cognitive profile of people with little or no sensory imagery, many of whom nonetheless create) is a useful caution that the engine has more than one implementation: recombination does not require pictures in the head.

The empirical evidence that this machinery does real creative work comes from three directions. Green and colleagues (finding 4) show far analogical mapping is a measurable neural operation. Dunbar’s ethnographies of world-class molecular biology labs found analogy in constant, load-bearing use in live discovery — the most productive labs reasoning through structured mappings from adjacent domains, week in, week out. And at civilisational scale, Uzzi’s atypical-combinations result and Simonton’s combinatorial models of scientific creativity describe discovery itself as constrained recombination — the same operation blending theory describes at the level of a single thought.

What the evidence does not show: that artists possess a different analogical mechanism from everyone else — the machinery is universal, which is rather the point — and blending theory in particular is more descriptive framework than falsifiable prediction; it earns its place by organising evidence, not by surviving crucial tests. The artist-specific claim is about usage, range and deliberateness: art is the practice of making blends as the product itself, every metaphor a shipped unit of combinatorial thought, whereas most professions treat the operation as incidental.

The “so what”: treat analogy as a discipline, not a decoration. Maintain live, deep source domains far from your field — the Root-Bernstein avocation evidence (finding 10) is arguably this mechanism wearing biographical clothes. In teams, make structured analogising explicit practice: state the mapping, push it past surface resemblance to relational structure (Gentner’s criterion), and ask what emergent structure the blend contains that neither input had. That last question is, mechanically, where new things come from.

8. The original personality exists — and it is measurable

There is a stable, measurable personality signature of original people — openness to experience above all — and it splits revealingly between artistic and scientific forms.

The anchor is Feist’s 1998 meta-analysis of personality in scientific and artistic creativity — still the reference synthesis — which found openness to experience the strongest and most consistent personality correlate of creative achievement, alongside a recognisable profile: autonomy, dominance of one’s own judgement, low conventionality. Its most useful result is comparative: creative artists and creative scientists share the openness core but diverge around it — artists higher on affective instability and norm-rejection, scientists on conscientious drive — not two unrelated temperaments but one trait engine with two exhausts. The lineage runs back to Barron’s classic 1955 study of the disposition toward originality, which identified the signature directly: original individuals prefer complexity over simplicity and hold to independence of judgement under group pressure — a disposition, note, describable as values and habits, not a mystery of birth.

Modern work has sharpened the construct. Kaufman’s four-factor analysis opens openness up into distinguishable appetites — including the split between experiential-aesthetic engagement and intellectual engagement — and his later study with colleagues delivers the clean double dissociation: openness predicts creative achievement in the arts; intellect predicts it in the sciences. Beaty and colleagues link openness to functional connectivity of the default network — the personality trait touching the same neural machinery as idea generation itself (findings 3 and 4). Carson, Peterson and Higgins’ Creative Achievement Questionnaire — developed inside the Harvard research programme that tied creative achievement to reduced latent inhibition, the loosened filtering of nominally “irrelevant” stimuli — gave the field its standard instrument for real-world creative attainment. Araki’s treatment of polymathy, with Alabbasi and Runco’s study of cross-domain creative activity in gifted students, adds breadth itself as a measurable disposition, and the experience-sampling study of mood and everyday creativity in the library shows the disposition operating in daily life — little-c creative activity woven through ordinary days, tracked in real time rather than reconstructed in retrospect. And on the perennial madness question, Kyaga and colleagues’ Swedish registry study of 300,000 people is the sober corrective: the creativity–psychopathology association is real but modest, familial and disorder-specific — closer to “shared familial cognitive style, in relatives more than patients” than to the romantic mad-genius equation.

What the evidence does not show: causation or fixity. Personality here is correlational; openness scores rise with experience and training as well as predicting them, and nothing in this literature licenses hiring by questionnaire or writing anyone off. Feist’s profiles describe distributions with wide overlap, not types.

The “so what”: stop selecting for polish and calling it talent. If openness plus independence of judgement is the measurable signature of original people, most institutional filters — consensus interviews, conformity-rewarding review, penalty for odd trajectories — are tuned to screen it out; the bias-against-novelty result from the science angle (Wang, Veugelers and Stephan) is the same filter operating on papers instead of people. Practical translation: weight evidence of self-directed cross-domain work (the CAQ logic) over credential smoothness, protect the complexity-preferring dissenter Barron described, and treat your own openness as trainable surface area — new domains, new mediums, new company.

9. Incubation, walking, embodiment — the body is part of the thinking

Stepping away from the problem measurably helps solve it, walking measurably lifts idea generation, and skilled creators demonstrably think with their bodies and their rooms — cognition does not stop at the skull.

The strongest single result is Sio and Ormerod’s Psychological Bulletin meta-analysis of incubation: across the experimental literature, setting a problem aside genuinely improves later solving — a positive overall effect, strongest for divergent-thinking tasks, exactly where originality lives. Gilhooly’s account supplies the mechanism candidates: continued unconscious associative work and the release from fixation — stepping away lets the wrong frame die. Oppezzo and Schwartz’s Stanford studies gave embodiment its most famous datum: across four experiments, walking substantially boosted creative ideation, indoors or out, with the effect persisting briefly after sitting back down — among the cheapest reliable creativity interventions on record.

The artists extend the claim from “breaks help” to “the body computes”. Kirsh’s study of professional dance choreography documents marking — dancers sketching movements with their bodies at low amplitude — as genuine physical thinking: the body used as a computational medium in which options are generated and tested more effectively than by mental simulation alone, the bodily counterpart of Tversky’s sketch results (finding 6). Malinin’s review frames this in 4E terms — cognition as embodied, embedded, enacted and extended — with the studio itself as part of the thinking system, which is precisely how Glăveanu’s craft-ethnography work (angle 3) describes working artists: intelligence distributed across hands, tools and workshop. Bateson’s essay on play supplies the deep lineage: playful states generate novel behavioural combinations across species — play is evolution’s own generative mode, and the improvising artist (finding 5) its adult professional. The social body counts too: Sawyer and DeZutter’s studies of improvised theatre document collaborative emergence — group creations arising from interaction that no member individually holds or could have produced — with Pels and colleagues’ scoping review mapping the young “group flow” literature that surrounds it, and Barrett, Creech and Zhukov’s systematic review confirming how much professional artistic creation is collaborative in practice. Literat and Glăveanu extend the same logic to the internet age, showing how genuinely distributed creative processes work when creation is spread across many contributors who never share a room. Csikszentmihalyi’s flow research frames the state side: deep absorption with immediate feedback as the signature experience of skilled creative work.

What the evidence does not show: incubation effects are modest and condition-dependent, not magic sleep-on-it alchemy; group flow remains conceptually loose (Pels’ review is candid about definitional sprawl); and 4E creativity is stronger as framework than as tested prediction. Walking is the exception — a clean, replicated experimental effect.

The “so what”: the working day is a cognitive instrument, currently mistuned. The evidence justifies specific design: hard problems deliberately interleaved with incubation gaps rather than ground through; walking meetings for generative agendas (and seated ones for evaluative agendas — finding 3); externalise early thinking into sketches, prototypes and physical mock-ups because hands find what heads miss; and stage genuinely interactive sessions where emergence can happen, rather than serial monologues around a table. None of this is wellness garnish. It is process engineering on meta-analytic ground.

10. Artistic originality transfers — the Root-Bernstein evidence and its limits

The people who reach the top of science are many times likelier to sustain serious artistic practices — the strongest signal in the library, and the one whose causality is weakest.

The headline evidence is Root-Bernstein and colleagues’ study of Nobel laureates: compared with average scientists, laureates are many times likelier to maintain serious arts and crafts avocations — sustained adult practices, not childhood piano. Their 2019 PNAS-line follow-up widens the base: across STEMM professionals, arts, crafts and design avocations track scientific achievement, and the practitioners themselves describe their art as supplying tools for thinking — observation, visualisation, pattern, manipulation — the very stack of findings 1 through 9. The library’s essay on Ramón y Cajal gives the mechanism a face: his draughtsmanship was not decoration of his neuroscience but part of its method. The first-person bioart account in the transfer angle — scientists reporting that collaboration with artists changed their scientific thinking — and the PNAS colloquium paper on art-science exchange add contemporary texture. At the population level, Catterall’s NEA report finds arts-engaged low-income youth outperforming across four longitudinal databases, with Fiske’s Champions of Change compendium and the President’s Committee’s Reinvesting in Arts Education review assembling the broader education record; the NEA’s How Creativity Works in the Brain report shows the policy world itself convening neuroscientists to put this evidence base under arts funding.

Then the honesty. The OECD’s Art for Art’s Sake — the definitive critical review, and this report’s designated sceptic — went through the transfer literature and found the causal case for far transfer weak: most studies are correlational, selection effects are everywhere (who chooses art?), and claimed academic spillovers mostly fail rigorous tests. Schellenberg’s randomised trial — music lessons producing a small IQ gain in children — anchors the debate precisely because it is both positive and modest: the flagship causal result in music transfer is real but small, and later scepticism about grander claims flowed from exactly this line of work. Nobel-laureate avocations admit a selection story too: perhaps polymathic energy (finding 8) causes both the science and the painting. Rank ten reflects this squarely — the largest effects in the library, the thinnest causal warrant.

Yet the sober reading is not “nothing transfers”. It is that transfer is real where it is specific: the OECD review itself credits arts training with skills internal to its stack — and the causal wins in this library are precisely stack-shaped: improv training raising divergent thinking (Hainselin), drawing retraining perception (Schlegel; Fan and colleagues), walking lifting ideation (Oppezzo and Schwartz). What fails tests is the lazy claim — art class raises maths scores. What survives is this report’s claim: artistic practice trains identifiable cognitive operations, and those operations are the ones eminent scientists disproportionately possess.

The “so what”: defend arts practice with the honest argument, because the dishonest one has been audited and lost. For individuals: choose one serious artistic practice as deliberate cognitive training. For institutions and states — a mid-sized European country like the Czech Republic included — fund arts education as originality infrastructure rather than as cultural decoration, and evaluate it on what it demonstrably trains: perception, problem finding, generative fluency, the tolerance for unfinished problems. A curriculum audited against findings 1 through 9 is defensible in front of any finance ministry; a curriculum defended by maths-score spillovers is not. Measured the honest way, the investment case survives its sceptics — and the country that acts on it early is buying the input the machine era makes scarce.

What to do first

The evidence assembled here supports a short, unromantic instruction set. First, pick up a practice, not a theory: one artistic discipline pursued seriously — drawing, an instrument, improv — is the validated delivery mechanism for the stack, retraining perception (finding 1), rehearsing the generative state (finding 5), and exercising the combinatorial engines (finding 7) in one weekly commitment. The Nobel evidence says the best scientists already do this; the training studies say it is not too late to start.

Second, re-engineer one process. Separate generation from evaluation in your team’s calendar (finding 3); budget problem-finding time before any solution work (finding 6); put walks and incubation gaps where the hard thinking is (finding 9); persist past the first wave of ideas (finding 4). None of this requires budget — only the authority to schedule.

Third, fix one filter. Audit whatever gate you control — hiring, funding, review — against the bias-against-novelty result and the openness evidence (finding 8), and change the single criterion that most punishes originality.

The deeper argument of this report is the reframe it opened with: originality is a capability — enumerable, measurable, trainable — and artists are its longest-running training tradition. What is required to build that capability deliberately, and how to run it as a system in the age of machine generation, are the questions of the two reports that follow this one. The evidence here settles the prior question: there is something real to build.

Written by the ENSI Foresight Division on two downloaded research libraries — “AI for Teaching at CTU” (107 primary documents) and “Original Thinking” (175 primary documents) — together with the Learning Exposure Index from ENSI’s “Effectiveness of Learning” project. The fourth report in the series that began with What Actually Works, The ČVUT Playbook and The Agentic University.

The argument, before the list

A randomised trial in Harvard’s introductory physics course, published in 2025, compared two ways of learning the same material. One was taught by expert instructors using active learning — the method that decades of education research place above almost anything else a lecturer can do. The other was an AI tutor built on GPT-4 and explicitly forbidden to hand out answers. Students with the tutor learned more than twice as much, in less time. Harvard’s CS50 has run a teaching agent for roughly 211,000 students at $1.50 per student per year. Georgia Tech’s Jill Watson reached over 96% question coverage at over 86% precision and saved its teachers more than 500 hours.

Read those three facts together and one conclusion is hard to avoid. The function universities were physically built around — explanation, delivered in a room, by an expert, at a scheduled hour — no longer belongs to them. It can be delivered at midnight, at any pace, for the price of a coffee per student per year, and in a well-specified domain it can outperform good human teaching. That does not make lecturers obsolete. It makes explanation a weak answer to the question every university now has to answer out loud: what do we produce that a student could not get more cheaply somewhere else?

The previous report in this series, The Agentic University, specified the machine that will run much of teaching at a technical university — eight agent archetypes, four data strata, four named human roles. Its closing argument was that the machine exists to buy back scarce human hours: the hours in which an experienced engineer sits with a stuck student and asks the question that reorganises their understanding. It did not say what those recovered hours should mostly be for. This report answers that, and the answer is its title. They should be spent producing original thinkers, because that is the last thing a university produces that the machines around it cannot.

The case rests on three findings that sit side by side in the libraries and are almost never read together.

First, the average has been automated. GPT-4 now outscores human norms on the classic divergent-thinking tests (Hubert, Awa and Zabelina). The best humans still win (Koivisto and Grassini), and professional writers still out-create language models on Torrance-style evaluations (Chakrabarty and colleagues). The middle of the human distribution has been matched; the top tail has not. At work, the same technology compresses the gap between beginner and expert — +34% for novices and almost nothing for experts in the NBER field study of 5,179 support agents. A graduate whose value lies in competent, conventional output is being priced toward the machine that produces competent, conventional output.

Second, the default use of AI makes people think alike. Doshi and Hauser found that AI-generated ideas make individual stories more creative while making the pool of stories more similar. Anderson and colleagues found people brainstorming with ChatGPT drifting toward the same ideas. Padmakumar and He found that co-writing with an instruction-tuned model narrows the diversity of what people write. In education the effect has a sharper edge. In the Bastani field experiment published in PNAS, roughly a thousand Turkish high-school students given raw GPT-4 access improved their practice grades by 48% — and then performed 17% worse than never-exposed peers once access was removed. The Microsoft Research and Carnegie Mellon study of 319 knowledge workers found that the more people trust the AI, the less critical thinking they do. Left on its defaults, an AI-saturated university will graduate people who are more fluent, more alike, and less practised at thinking on their own.

Third, originality can be trained. Scott, Leritz and Mumford’s meta-analysis of 70 studies found that well-designed creativity training reliably works. The OECD has run creativity pedagogy through classroom trials in eleven countries, and PISA 2022 assessed the creative thinking of fifteen-year-olds in 64. ENSI’s Original Thinking library decomposes originality into eight disciplines, each an operation that can be practised badly or well. None of them requires talent as an entry fee.

Put the three together and the strategic position is short. A university that runs on agents and does not deliberately produce original thinkers will become the most efficient producer of average graduates in its history. The opportunity is the mirror image of the risk. The hours the agents free up are exactly the hours in which originality is trained — because originality is trained the way the studio has always trained it: by making things, putting them in front of people who will tell the truth, and doing it again.

Two cautions, because this argument is easy to hear wrongly. It is not an argument against foundations. The Agentic University lists unassisted core reasoning as a graduate outcome for a reason: you cannot check what you could never have derived, and you cannot originate in a field you cannot reason in. At a university where 31.8% of first-year bachelor students failed in 2024 — 51.1% at the Faculty of Mechanical Engineering — the floor comes first. The claim is narrower: foundations are the floor; original thought is the product. Nor is it an argument that originality replaces certification. A degree remains a statement one institution makes to strangers about a person. The argument is that the statement must now include one more line: this person can find a problem worth solving, and can defend what they think about it.

“Monopoly” needs a qualification of its own. It is not permanent, and it is not owned by universities. It belongs to the top tail of human thinking, which moves every time the models improve, and a university only shares in it if it actually trains people to reach that tail. On present evidence most do not. In ENSI’s Learning Exposure scoring of 100 human life paths, formal education is the lowest-scoring of ten categories, and the elite research-university degree ranks 79th — held back above all by the fact that for three or four years almost no verdict on a student’s work comes from anything other than a marker applying a rubric. The monopoly is available. It has to be earned.

The strategy in brief

• Declare original thought the university’s product, with foundations as its non-negotiable floor — and add a fifth graduate outcome, origination, to the four specified in The Agentic University.

• Spend the hours the agents recover on the studio — small-group making, critique and defence — and commit to that in writing before the savings arrive (Playbook idea 14).

• Build the infrastructure of thinking on purpose, in seven elements: peers, public judgement, mentors who ask, real problems, teachers who are rewarded, protected silence, and the nerve to stand alone.

• Configure the AI to widen thinking rather than narrow it: humans generate first, agents critique afterwards, provenance is always tagged, and drift toward the average is measured.

• Assess own thought by defence, not by style — two-lane assessment, a Defence Week for ideas, final projects on problems the student found.

• Change what the institution rewards — evidenced teaching redesign counts for promotion, and every faculty has a named studio lead.

• Measure with a small portfolio of weak gauges read together, never a single target.

• Run the whole programme as research: every studio pilot a pre-registered trial, so the university ends up owning evidence instead of a slogan.

How this report is organised

Fourteen sections move from diagnosis to design to delivery. Sections 1–3 establish why original thought has become the product and why the agentic university will not produce it by default. Sections 4–6 establish that originality is a trainable practice, which of its disciplines matter most now, and why ideas must be experienced in an environment that tells the truth. Sections 7–8 specify the infrastructure, human and agentic. Sections 9–12 turn it into institutional machinery: the graduate outcome, assessment, incentives and measurement. Section 13 is the thirty-six-month sequence; section 14 names the ways it fails. Every study cited is in one of the ENSI libraries named above. Learning Exposure figures are calibrated expert-judgement scores, not psychometric measurements, and are labelled as such wherever they appear.

1. Explanation has been automated — and it was never the whole product

The best trials in the library show a constrained AI tutor matching or beating good human explanation at negligible cost. That ends the lecture’s monopoly. It does not end the university’s.

The evidence is stronger than most academic discussions of AI assume. Kestin, Miller and colleagues’ crossover trial in Harvard’s PS2 physics course, published in Nature Scientific Reports, pitted an AI tutor against expert-taught active learning — not against a bad lecture, but against the best-evidenced form of undergraduate physics teaching — and the AI condition won on a pre-specified learning measure. The World Bank’s six-week trial of GPT-4 tutoring in Nigeria returned 0.31 standard deviations, among the most cost-effective education interventions ever recorded. Stanford’s Tutor CoPilot trial, across 900 tutors and 1,800 students, raised topic mastery by 4 percentage points for about $20 per tutor per year. A meta-analysis of 35 experiments and 4,193 participants puts ChatGPT’s effect on learning at g = 0.670.

Scale and cost are no longer speculative either. CS50’s duck has processed around 10 million queries, and 94% of its students found it helpful. Jill Watson’s LLM-era rebuild answers course questions with 76.7% accuracy against 31.3% for a generic assistant — the grounding in course material, not the model, doing the work.

The limits matter and should be stated before anything is built on these numbers. Almost every trial measures learning at or near the end of a short intervention; the library contains no strong evidence on retention months later. Every successful system was built to withhold — Kestin’s tutor revealed one step at a time and refused to give the final answer. And the one deployment that published its failure rate, CS50, found that 22% of responses contained code despite instructions not to, because long conversations erode the guardrail. Explanation has been automated; good explanation has been automated only where somebody engineered it carefully and staffed the review loop.

Even with those limits, the direction is settled, and it forces a question universities have been able to avoid for eight centuries. The lecture hall was a solution to the scarcity of explanation. When explanation stops being scarce, the institution has to say what it was really selling.

Computing education supplies the cleanest answer, because it met the disruption first. Becker and colleagues titled their SIGCSE paper Programming Is Hard — Or At Least It Used To Be, and their argument generalises well beyond code. Many introductory learning objectives were proxies. Nobody actually wanted students to write a loop from memory; the discipline wanted them to decompose a problem, and writing the loop was how it checked. The proxy broke. The objective did not. The same is true of a whole degree. Recall of content was always a proxy for the ability to think inside a field — and thinking inside a field, at its best, includes thinking something the field has not yet thought.

That is the product the lecture was a delivery mechanism for. It is still scarce. It is the subject of the rest of this report.

2. The average graduate is being priced toward the machine

AI raises the floor and flattens the middle. The graduate whose value is competent, conventional work is competing with a system that does competent, conventional work for almost nothing; the graduate whose value sits in the tail is not.

Three lines of evidence converge. The first is head-to-head creativity testing. Hubert, Awa and Zabelina found GPT-4 outscoring human norms on standard divergent-thinking tasks — the average person no longer wins the average originality test. Koivisto and Grassini found chatbots beating average humans on the Alternate Uses Task while the best humans still outperformed them. Chakrabarty and colleagues found LLM-written stories passing fewer Torrance-style creativity tests than professional writers’ work. Read through Margaret Boden’s typology, the computational-creativity literature in the library points the same way: machines are strong at combinational and exploratory creativity inside a given space and weak at transformational creativity — changing the space itself.

The second is labour-market evidence. Brynjolfsson, Li and Raymond’s NBER study of 5,179 support agents found an average productivity gain of 14%, concentrated almost entirely in novices (+34%) with near-zero effect for experts, because the model diffuses the tacit knowledge of the best performers to everyone else. Dell’Acqua, Mollick and Lakhani’s field experiment on 758 BCG consultants found work inside the AI frontier rated 40% higher in quality — and, on a task just outside it, consultants 19 percentage points less likely to reach the correct answer than colleagues without AI. Developers with Copilot finished a standard task about 56% faster in the GitHub–Microsoft–MIT trial. And Stanford’s Digital Economy Lab, tracking payroll data, finds employment falling specifically for young workers in the occupations most exposed to AI — the entry-level technical roles a technical university’s graduates walk into.

The third is where employers and forecasters now place value. The World Economic Forum’s Future of Jobs Report 2025 ranks creative thinking among the fastest-rising core skills, precisely because routine cognition is being automated faster than the generation of new framings. Nesta’s Creativity vs Robots analysis reached the same structural conclusion years earlier: creative occupations are the ones that resist substitution.

Read together, the picture is uncomfortable for any university whose graduates are defined mainly by what they know. Compression means a CTU graduate may perform like a competent junior on day one and find no junior role in which to become a senior — the “hollowed apprenticeship” that The Agentic University named as the one failure mode nobody has solved. Compression also means that the signal value of ordinary competence shrinks: when everyone with a model can produce the median answer, the median answer stops distinguishing anyone. What remains scarce is the thing the jagged-frontier experiment shows people cannot do untaught — knowing where the machine is wrong — and the thing the creativity studies show the machine does least reliably: the framing nobody holds, the combination nobody has licensed.

Two honest limits. The compression studies come from work settings, not from engineering degrees, and they measure task performance rather than durable expertise. And “the tail” is a moving target: every model improvement redefines what counts as beyond the machine. Neither limit weakens the strategic conclusion. If the target moves, the capability that matters is the one that keeps moving with it — which is a practice, not a stock of knowledge.

3. Left alone, the agentic university manufactures convergence

Every convergence result in the library runs through the same mechanism: AI entering the thinking process before the human has formed their own view. An agentic university that does not control that sequence will make its students more alike.

The individual-level evidence is now replicated across settings. Doshi and Hauser’s experiment showed generative-AI ideas raising the creativity of individual stories while shrinking the collective diversity of the pool — everyone slightly better, everyone much more similar. Anderson and colleagues found the same convergence in live ideation with ChatGPT. Padmakumar and He found that merely co-writing with an instruction-tuned model reduces the diversity of the human-written text itself. None of these studies required anyone to cheat. They describe ordinary, well-intentioned use.

Education adds a second mechanism: the removal of productive struggle. In the Bastani PNAS experiment, unrestricted GPT-4 access raised practice performance by 48% — 127% for a tutor-prompted version — while the unrestricted group performed 17% worse than controls once the tool was withdrawn. The difficulty had been removed, and the removal felt like progress. Teacher-designed hint scaffolds largely prevented the loss, which is the point: the design around the model decides the outcome. Prather and colleagues, watching first-year programmers use Copilot, documented drift — the student’s mental model silently diverging from the code accumulating on screen — alongside over-trust and a collapse of the metacognitive loop. Lee and colleagues’ Microsoft–CMU study found that confidence in AI predicts less critical-thinking effort and that effort shifts from producing judgement to checking someone else’s. Students sense this: in MIT’s 2026 survey of 1,002 affiliates, 90% were concerned about overreliance, 67% very concerned.

Now add the property of deployed agents that The Agentic University documented. CS50’s duck was built to refuse solutions and still handed out code in 22% of responses and 48% of conversations, because in a long exchange the system prompt loses authority and the model reverts to being maximally helpful. An agent’s natural pull is toward giving the answer. Scaled across a university, that pull is a pull toward the centre of the distribution — the most probable framing of every question, delivered first, to every student.

The complementarity literature closes off the easy hope that human plus machine automatically beats either. Hemmer and colleagues’ review finds the empirical record frequently disappointing: many human–AI teams underperform the better of their two members. Complementarity is engineered, and it is engineered mainly by deciding where the boundary sits.

The conclusion is not that the agentic university should ban AI from thinking work. That would forfeit the tutoring gains in section 1 and train students for a profession that no longer exists. The conclusion is about sequence. The convergence effects arise when the model enters the generative phase — before the student has a view. Criticism of a finished human draft is a categorically different exposure from suggestion during composition. The same model that homogenises a brainstorm can sharpen a finished argument. An agentic university therefore needs a layer where the usual rules are inverted: explanation agents answer; studio agents only question, and only afterwards. Section 8 specifies that layer.

4. Originality is a practice, not a gift — and it can be taught to everyone

The evidence treats original thinking as a set of trainable operations, not a temperament. That makes it a curriculum question — and a question for every student, not an honours track.

Start with the definition, because it already rules out the two popular misreadings. The field’s standard definition, fixed by Runco and Jaeger, makes creativity a conjunction: originality and effectiveness. New alone is noise; useful alone is a textbook. Original thinking is the discipline of producing things that are both, and the conjunction is exactly why it is hard — novelty pulls away from what works, effectiveness pulls back toward what exists. That definition suits an engineering school unusually well. An engineer’s original idea has to run.

The trainability evidence is broad. Scott, Leritz and Mumford’s meta-analysis of 70 studies found well-designed creativity training reliably effective, where “well-designed” means grounded in cognitive mechanism and realistic practice rather than inspirational theatre. Epstein decomposed the trainable core into four measurable competencies — capturing new ideas as they occur, challenging oneself with hard tasks, broadening one’s repertoire, and surrounding oneself with varied stimuli. Hainselin and colleagues raised teenagers’ divergent-thinking originality with an eleven-week improvisation course. Schlegel and colleagues found that art training measurably changes neural structure and function over months. At system scale, the OECD’s Fostering Students’ Creativity and Critical Thinking programme ran rubric-based creativity teaching through classroom trials in eleven countries, and PISA 2022 assessed creative thinking across 64. Valgeirsdottir and Onarheim’s review of realistic creativity training adds the design constraint that matters most for a university: training sticks when it is embedded in real work, not delivered as an off-site exercise.

ENSI’s Original Thinking Framework organises the evidence into eight disciplines, derived from the one profession whose entire output is originality — artists — and checked against the science and entrepreneurship literatures for transfer:

• The Trained Eye — perceiving past your own categories; most people misperceive before they mis-think.

• The Found Problem — deciding what the problem is before solving it.

• The Long Reach — connecting semantically distant material, deliberately and past the obvious first ideas.

• The Blending Engine — carrying structure from one domain into another to produce something neither contained.

• The Oscillation — separating generating from judging, and switching between them on purpose.

• The Loved Constraint — using limits to block clichés.

• The Nerve — the disposition to stand alone while original work is punished before it is rewarded.

• The Thinking Hand — making things early, because the made thing is where the thinking happens.

Two features of the framework make it institutionally usable. It is mechanistic: each discipline names an operation, not a virtue. And, following Kaufman and Beghetto’s Four C model, the same disciplines operate at every level — the student’s first genuine insight, the hobbyist’s project, the professional’s contribution — differing in load rather than in kind. That second feature settles a strategic choice. Original thinking is not an elite track for the top five per cent. It is a practice every student can run, at their own level, from the first semester.

The transfer evidence also argues for including the arts rather than treating them as decoration. Root-Bernstein and colleagues found Nobel laureates far more likely than average scientists to keep serious arts and crafts avocations, and a later PNAS study found the same association across STEMM professionals. The honest limit belongs next to it: the OECD’s Art for Art’s Sake review found the popular claim that arts education makes people generally smarter weakly supported. The case here does not rest on that claim. It rests on specific practices — observation, problem construction, constraint work, making — that are trained hardest in studios and are individually evidenced in their target domains.

The implication for a technical university is direct. Originality is not something to hope students bring with them or discover on their own. It is a set of repetitions the institution either schedules or does not.

5. The three disciplines machines handle worst are the three universities teach least

The Original Thinking Framework’s reading of the head-to-head evidence is that models are strongest in the middle range of association and weakest at the Found Problem, the Nerve and the Thinking Hand. Those three are precisely what a conventional degree gives students almost no practice in.

The Found Problem. Reiter-Palmon and Murugavel’s review establishes that how a person constructs the problem shapes both the process and the creativity of the result; people who spend deliberate effort re-representing an ill-defined situation produce more original work. Botella, Zenasni and Lubart’s study of art students found the creative process front-loaded with the artist’s own definitional work — deciding what the piece is even about. Scotney and colleagues found cross-domain inspiration strongest in exactly this early, problem-finding phase. Now compare the typical engineering degree. From the first problem set to the final exam, the problem arrives already framed, with its boundary conditions, its method and often its answer format specified. Students get thousands of repetitions of solving and almost none of finding. The framework’s behavioural test is simple and damning: an original thinker can tell you, for any project, which problem they rejected and why. Most graduates have never rejected a problem in their academic lives, because none was ever theirs to reject.

This matters doubly in the AI era. A model asked a question returns the most statistically probable framing of it — the framing everyone else also receives. Foster, Rzhetsky and Evans found scientists’ research strategies clustering around tradition because the reward system punishes the variance of risky innovation. A student who has never framed their own problem will accept the model’s framing, and the field’s, by default.

The Nerve. Wang, Veugelers and Stephan’s NBER work documents the bias against novelty in science: novel papers suffer delayed recognition and bibliometric disadvantage despite higher long-run impact. Sternberg and Lubart’s investment theory describes what original people actually do — buy low and sell high in the world of ideas, holding unfashionable positions until the field catches up — and that strategy only works for someone who can tolerate the holding period. Feist’s meta-analysis finds openness the strongest personality correlate of creativity, with independence and nonconformity marking creative scientists and artists alike; Barron’s classic work finds original people preferring complexity and judging independently of the room. Azoulay’s evidence from long-horizon HHMI funding shows the institutional version: tolerating early failure causally increases breakthroughs. Universities, by contrast, are built to grade: a strange answer that turns out wrong costs marks, and a strange answer that turns out right often costs marks too, because the rubric did not anticipate it. The framework is explicit that the Nerve is trained socially — through regular, survivable doses of public judgement, as in the studio crit — and it adds a caution the institution must keep: Kyaga’s Swedish registry study of 300,000 people shows the open, independent, “leaky-filter” configuration sits close to real vulnerabilities. The Nerve needs scaffolding, not romance.

The Thinking Hand. Tversky’s sketch-cognition research shows designers discovering relations in their own sketches that they did not knowingly put there. Kirsh found choreographers thinking physically by “marking” movements, outperforming pure mental simulation. Oppezzo and Schwartz showed across four experiments that walking alone boosts creative ideation. Making is not the record of thought; it is part of the thought. Engineering education should own this discipline outright — its native forms are the lab, the design studio and the build. Yet What Actually Works documents three decades of drift away from artefacts toward the cheap end of assessment: the individually submitted problem set, the templated lab report, the multiple-choice test, each of which a model now completes in seconds.

In Boden’s terms, these three are why the transformational end resists automation. Finding a problem, holding a position against the room, and learning from what a physical thing does when it is built are all transformational moves — they change the space rather than search inside it. They depend on stakes, embodiment and a willingness to be visibly wrong. Those are exactly the conditions a model does not have and a conventional degree does not supply. The strategic consequence is precise: train hardest where the machines are weakest and where the curriculum is currently thinnest.

6. Ideas have to be experienced — in an environment that tells the truth

Original thinking is not transmitted; it is formed by attempting, being wrong, and finding out quickly and honestly. ENSI’s Learning Exposure scoring suggests universities are unusually poor at the “finding out” part — and that fixing it is a design choice, not a question of effort.

The Learning Exposure Index, built for ENSI’s Effectiveness of Learning project, scores 100 durable human paths on 48 dimensions. Its scores are calibrated expert judgements against explicit anchors rather than validated psychometrics, and they should be read that way. Its most important result is a correlation of -0.00 between two of its families: Heart — how much of a person’s identity, morals and emotions a path engages — and Feedback Ecology — whether the environment can tell the person the truth about their performance. How meaningful an experience feels says nothing about whether it teaches. Meaning is not evidence.

The construct underneath is Hogarth’s distinction between kind learning environments, which return fast, accurate signals, and wicked ones, which return slow, noisy or misleading signals — or excellent signals about the wrong thing. Kind environments build real expertise. Wicked ones build confidence without competence.

Scored against that distinction, formal education comes out badly. It is the lowest-scoring of ten categories, at a mean index of 42.3; creative practice is the highest at 62.1. The elite research-university degree ranks 79th of 100 and the mass-market university 93rd. The mechanism is specific. The elite-university dossier scores reality contact at 1: for three or four years, every verdict on a student’s work comes from an intermediary with a rubric — never a user, a client, an opponent or a physical system. Feedback fidelity and latency score 2; mastery legibility 2, meaning a student can finish a famous degree without knowing how good they are. Consequence weight is 1 while identity entanglement is 3 — a high stake in the self sitting on top of a minimal stake in the world.

Independent measurement points the same way, with a necessary counterweight. Arum, Roksa and Cho’s longitudinal study found gains of only about 0.18 standard deviations in critical thinking, complex reasoning and writing over the first two years of college. Mountjoy and Hickman found that selectivity barely predicts a college’s value-added. But Ritchie and Tucker-Drob’s meta-analysis finds each additional year of education raising measured cognitive ability by roughly 1 to 5 IQ points, and the earnings premium is real. The claim is not that university fails to build capability. It is that universities convert a small share of a student’s years into the kind of experience that forms independent judgement — because the verdict almost never comes from reality.

The learning-science literature says what a better environment looks like, and it contains a distinction that is routinely collapsed. Bjork and Bjork’s desirable difficulties — spacing, interleaving, retrieval, generation — make learning feel harder and produce more durable results; Roediger and Karpicke found tested material recalled at 61% against 40% for restudied material after a week. Kapur’s productive-failure studies found students who attempted ill-structured problems before instruction outperforming directly-instructed peers on conceptual understanding and transfer. Metcalfe found errorful generation followed by correction beats errorless study. Every one of these is difficulty with feedback attached. Difficulty without feedback is not desirable; it is opacity. Productive failure is productive only because instruction follows it.

Two more dimensions carry the design. The first is error affordability — whether being wrong is cheap enough to repeat. It is the only one of the index’s 48 dimensions that correlates negatively with total exposure (-0.13): demanding environments tend to make failure expensive. The inversions are instructive. Stand-up comedy scores error affordability at the maximum and feedback fidelity at the maximum — laughter is involuntary and arrives within a second — and ranks joint second of all 100 paths. Commercial aviation decoupled consequence from cost by building the simulator, and McGaghie and colleagues’ synthesis shows the medical equivalent, simulation-based mastery learning, improving real patient outcomes. The second is deliberate-practice affordance — whether the hard part can be isolated and repeated. Macnamara and colleagues’ meta-analysis of 88 studies found practice explaining about 26% of performance variance in games but under 1% in professions, a gradient that tracks how structured the environment is rather than how hard people work. Deans for Impact’s summary is blunt: most experience does not produce expertise unless it is deliberately structured.

There is a warning inside the same data. Studio visual art has the highest “grip” score in the catalogue — the most absorbing, most personally engaging path — alongside a feedback ecology of only 41.7. An environment can hold a person completely for a decade and never once tell them the truth. A university studio built on the art-school model alone would reproduce that failure: intense, meaningful and uninformative. The model to copy is a hybrid — the studio’s making, the comedy club’s fast and honest audience, and the simulator’s cheap failure.

That hybrid is closer to hand at a technical university than anywhere else. What Actually Works calls engineering’s artefact-based pedagogy the structural advantage nobody is exploiting: a structure is checked by statics, a circuit oscillates or it does not, a robot finishes the task or falls over. The verdict is external to the examiner. So the design brief for “experiencing ideas” reduces to four levers the index isolates:

• Raise reality contact — let a user, a client, a physical system or a competitor deliver the verdict wherever possible.

• Make error cheap — simulators, digital twins, prototypes, rehearsal defences, repeated attempts with no grade attached.

• Shorten the loop — feedback in days, not at the end of the semester.

• Isolate the hard part — let students repeat problem framing, critique and defence as separate, practised skills.

None of those levers requires more hours from anyone or more rigour. They require the environment to be honest, on a schedule, at a cost the student can afford.

7. The infrastructure of thinking — seven elements

Independent thought does not appear on its own, and it does not come from lectures or well-meant advice. It forms in an environment that pushes back. That environment has seven parts, and at most universities each of them exists only by accident — usually because one teacher builds it in spite of the system.

A school has buildings, curricula and examinations. What it rarely has, by design, is the set of conditions under which a young person starts thinking for themselves and keeps doing it when that becomes costly. The seven elements below are that set. Each is specified the same way: what it does, what the evidence says, what it looks like at CTU, what the agents may do, and the boundary that must hold.

1. Peers who make you say what you mean

• What it does. An idea that never leaves your head is not yet an idea. Equals force it out — they ask what you actually claim, and they build on it or break it.

• The evidence. Sawyer and DeZutter’s work on collaborative emergence shows collective originality arising from visible contributions others can build on; nothing can be built on what stays private. Glăveanu’s craft-ecology fieldwork shows creative communities deciding, together, which departures from tradition count as contributions. Dunbar’s in-vivo studies of molecular-biology labs found discoveries emerging from distributed reasoning and analogy in ordinary lab meetings. Wu, Wang and Evans, across 65 million papers, patents and products, found that small teams disrupt and large teams develop.

• At CTU. Studio cohorts of roughly eight to twelve students, deliberately mixed across the eight faculties, running from the second year onward. Cross-faculty build projects of the kind the Playbook proposes (idea 20: FIT and FEL students building tutoring agents for FS and FSv courses, with the receiving academic as client).

• The agents. The Archivist keeps each cohort’s shared corpus of sketches, drafts and dead ends, with every item tagged by provenance.

• The boundary. No agent takes part in a group’s generative session. Groups present rough artefacts early; the polished pitch is banned in the first round, because collective originality needs something unfinished to build on.

2. Public judgement that finds your real limits

• What it does. Without comparison against others, a person never discovers where their real limits are — and that they usually sit further out than they thought.

• The evidence. The strongest evidence in the libraries concerns exposure rather than competition as such. Stand-up comedy — publicly attributed work, weekly rejection, an audience that cannot fake a laugh — scores at the maximum on feedback fidelity and error affordability and ranks joint second of 100 paths in the Learning Exposure scoring. The studio crit is described in the Original Thinking library as a deliberate exposure regime: regular, survivable doses of public judgement. The same scoring carries a warning: when the verdict comes from a panel applying its taste, the fidelity ceiling is low however sophisticated the panel.

• At CTU. Open-problem challenges with industry and public-sector partners, judged wherever possible by something that does not care about the entrant — a benchmark, a live user, a working prototype, an adversarial test. The university’s existing talent environments, such as FIKS at FIT and FEL Camp, are the seeds; the Playbook already proposes scaling them (idea 15).

• The agents. The Red-Team forecasts how judges and users will attack an entry, so the entrant walks in prepared rather than protected.

• The boundary. Winning must depend on what works, not on polish or presentation. A competition judged on slides trains slide-making.

3. Mentors who teach how to ask, not what to think

• What it does. A mentor does not supply the conclusion. They show the student how to question their own framing, and they are close enough to say, specifically, “this is wrong”.

• The evidence. Every AI tutor that produced large learning gains was built to withhold. Stanford’s Tutor CoPilot did not tutor students at all; it coached human tutors in real time, and helped students of the weakest tutors most (+9 percentage points). A 2025 systematic review of deliberate practice in psychotherapy training found coached, feedback-rich practice outperforming standard training even in open-ended, judgement-heavy work. In the Learning Exposure scoring, mentorship density scores 4 for airline pilots and emergency-medicine residents — and 1 for the mass-market university, whose dossier records “a striking absence of anyone who ever learned their name well enough to tell them they were wrong.”

• At CTU. Every student in the studio years has one named mentor who knows their work. Doctoral students at CIIRC and the faculties’ research groups are a natural mentor pool, and the Playbook’s proposal to make teaching-AI a doctoral topic (idea 26) extends to studio practice. Industry engineers mentor through the Junior Engineer Compact (idea 4).

• The agents. The Studio Critic — already specified in The Agentic University as “critique, never generation” — questions students’ work between sessions. A Tutor CoPilot–style assistant can coach the human mentor on what to ask.

• The boundary. Agents may ask questions; they may not supply a framing. The mentor’s job, human or machine, is to make the student’s thinking better, not to replace it with better thinking.

4. Real problems, where an idea meets reality and one of them gives way

• What it does. An idea tested only against a rubric has never been tested. Real problems return a verdict the student cannot argue with.

• The evidence. Reality contact is the dimension on which the elite university scores 1 and the one the Learning Exposure work calls the single highest-leverage change available. Emergency-medicine residency shows high stakes combined with fast, truthful feedback. Sarasvathy’s expert entrepreneurs structure attempts around affordable loss, so being wrong is survivable and therefore repeatable. Camuffo and colleagues’ randomised trial found founders trained to treat beliefs as testable hypotheses made measurably better pivot decisions. CESAER’s Engineer of the Future white paper and the CDIO tradition already commit European technical universities to challenge-based learning.

• At CTU. The Junior Engineer Compact — supervised responsibility for real work in the final year — and a pipeline of real problems through EDIH CTU, the AI-MATTERS testing facility and the university’s industrial partners. Strategic-plan goal 1.3, “bring practice into teaching”, is the existing mandate. Crucially, the problem arrives as a situation, not a specification: the student’s first deliverable is the problem statement.

• The agents. The Scout brings in mechanisms from distant fields; the Lab and Simulation Agent provides digital twins so that failure is cheap before it is expensive.

• The boundary. The partner gets useful work; CTU certifies the judgement. If assessment is of output alone, the scheme becomes free labour, which the Playbook names as an ethical failure mode.

5. Teachers who can bear all of this — including being surpassed

• What it does. Every element above is delivered by academics. Studio teaching is harder than lecturing: it requires tolerating ambiguity, critiquing without taking over, and being pleased when a student sees further than the teacher.

• The evidence. Ithaka S+R’s interviews across nineteen universities found teaching innovation happening in isolated pockets and dying on friction — no time, no recognition, nobody to ask. The multi-institution barriers study found those obstacles operating independently at individual, departmental and institutional level, so fixing one changes little. The Playbook rates changing promotion criteria as the highest-leverage idea on its list and notes that not one study has tested it.

• At CTU. 2,247 academic staff whose careers currently reward publication. Evidenced teaching redesign counts for promotion (idea 6); redesign fellowships give released time with a deliverable (idea 17); ETH Zurich’s lecturer framework is adopted rather than rewritten (idea 11), extended with studio skills such as running a critique.

• The agents. The Course Concierge and Instructional Design Agent absorb logistics and first drafts, which is where the teaching hours come from (section 8).

• The boundary. Studio teaching is credited, not added. Asking academics to run studios on top of their existing load is how the whole strategy fails quietly.

6. Silence, so that your own thought has a chance to speak

• What it does. Without periods in which nobody wants anything, an independent thought has no moment to surface. In an AI-saturated environment, silence has to be scheduled.

• The evidence. Sio and Ormerod’s meta-analysis in Psychological Bulletin found incubation effects real and positive, strongest for divergent problems; Gilhooly’s work supplies the mechanism of unconscious processing. Oppezzo and Schwartz found that walking boosts creative ideation. Ellamil and colleagues showed that generating and evaluating are different brain states, which is why the studio separates them. Torrance built incubation into his model of teaching decades ago. And the Bastani experiment shows what the absence of struggle costs.

• At CTU. Projects are opened in one session and returned to after a deliberate gap, rather than completed in one sitting. Generation sessions are AI-free by rule. The Playbook’s “cognitive gym” (idea 35) — deliberately unassisted practice, defended pedagogically rather than punitively — is the formal home of this element.

• The agents. None, by design.

• The boundary. Silence is scheduled support, not abandonment. The student is alone with the problem on purpose, for a defined period, with a mentor waiting at the end of it.

7. The nerve to think differently — and to withstand it

• What it does. Thinking independently is one skill. Holding the result is another. As soon as someone starts thinking their own way, the environment tends to push back — first with ridicule, then suspicion, then isolation — and the most able people learn to hide exactly what is most valuable about them.

• The evidence. Original work is punished before it is rewarded (Wang, Veugelers and Stephan). Original people buy low and sell high in ideas, which requires tolerating a holding period (Sternberg and Lubart). Independence of judgement and preference for complexity mark original people (Barron; Feist). Amabile’s componential theory makes intrinsic motivation central precisely because external pressure is what wears originality down. The Original Thinking Framework treats the Nerve as trained socially, through repeated and survivable exposure — and warns, via Kyaga’s registry data, that it needs scaffolding.

• At CTU. Public defences in which a well-argued position that turns out wrong can still score well; explicit reward for instructive failure; cohorts large enough that nobody holds an unpopular idea alone. The studio also teaches the most important distinction a young thinker can learn: substantive criticism improves an idea; social pressure only wants it silenced. A crit that separates judgement of the work from judgement of the person teaches that distinction by practice. Resilience is not trained in isolation; it is built in a community pulling in the same direction.

• The agents. The Sparring Partner produces the strongest case against a finished draft, so the student meets real opposition in a safe setting first. The Red-Team treats a predicted novelty penalty as information about timing and framing, never as a reason to stop.

• The boundary. Nobody is graded on agreement with the examiner. Examiner calibration (Playbook idea 19) is the safeguard.

8. Where the hours come from: the agentic university, re-pointed

The studio costs contact time. The agentic university produces contact time. The strategy works only if the second is committed to the first — and if the agents inside the studio are configured to widen thinking rather than narrow it.

The Agentic University specifies eight agent archetypes. Four of them are, in effect, hour-recovery machines:

• The Course Concierge absorbs logistics — deadlines, rules, the question asked forty times a semester.

• The Gateway Tutor absorbs repeated explanation in the courses where students most often fall behind.

• The Feedback Agent absorbs first-pass formative feedback, never summative marking.

• The Instructional Design Agent drafts syllabi, slides and exercises for human authorship.

Georgia Tech’s 500-plus saved teacher hours came largely from exactly this kind of work. The Playbook’s idea 14 names the choice that follows: recovered hours become either better teaching or a staffing cut, and that choice is the difference between an agentic university and an automated one. The rule this report proposes is the Playbook’s, made specific: before the savings materialise, commit in writing that recovered hours go to studio, critique, defence and mentoring. The same passage names the trap. The supervision loop that makes the agents work — the agent steward sampling conversations each week — is the line item cut first, and nothing visibly breaks for months.

Inside the studio, a different configuration applies. The Studio Critic already exists in the architecture. To it, this report adds a student-facing version of the six agents specified in ENSI’s Originality Engine, each with the rule that keeps it from homogenising:

• The Scout sends each student or cohort a weekly dispatch of mechanisms from fields they have never worked in, with a distance quota and no relevance ranking — ranking is a convergence operation.

• The Sparring Partner sees work only after a complete human draft exists. It criticises and never rewrites, and it always argues the strongest opposing case rather than a balanced review.

• The Divergence Auditor works for the programme, not the student’s grade. Each semester it measures how far theses and projects sit from the field’s mainstream, and how far they sit from each other, and flags cohorts where output rises while distance shrinks — the signature of creeping AI dependence. It is never visible while students are drafting, and it never becomes a mark.

• The Blender takes a problem stated as a structure — entities, relations, constraints — and returns analogies from remote domains, justified relation by relation. It returns mappings, not solutions.

• The Archivist maintains each student’s idea portfolio and tags every item as human-generated, AI-assisted or AI-produced. That tagging is what later makes it possible to assess own thought at all.

• The Red-Team forecasts how examiners, reviewers or industry partners will receive a novel idea, and where its novelty will be misjudged.

Five rules from the Originality Engine govern the layer, with one addition for a university:

• Humans generate first. No agent contributes content before a complete human first attempt exists.

• AI diverges; humans converge. When agents do produce options, they are asked for the unusual ones; selection and synthesis stay with the student.

• Never accept the model’s first framing. The student reframes at least once, unaided, before any agent proceeds.

• Measure drift every semester. Convergence is gradual and invisible from inside; it shows only in the numbers.

• Provenance or it did not happen. Every artefact carries its human/AI tag from the start.

• Students always know which layer they are in. The explanation layer answers questions; the studio layer only asks them. Mixing the two is how the tutor’s helpfulness leaks into the studio.

The same operating model applies as for every other agent. A named agent steward samples studio-agent behaviour, because instruction dilution will pull the Sparring Partner toward rewriting just as it pulled CS50’s duck toward handing out code. The evidence lead evaluates the studio agents like any other deployment. And two CTU-specific cautions from the Playbook carry over: the layer should be CTU-built and grounded (idea 9), with the interaction stream retained by the university (idea 21), and frontier models serve Czech technical language measurably worse than English (idea 32), so studio agents working in Czech need their own evaluation before they are trusted.

9. A fifth graduate outcome: origination

The four outcomes in The Agentic University are defensive — they protect graduates and their clients from machine error. The fifth is generative: it certifies that the graduate can supply what the machine cannot.

The Playbook’s idea 5 proposes that every CTU programme map four outcomes to named assessments: verification, frontier judgement, unassisted core reasoning, and accountability for results one did not personally generate. Each traces to specific evidence, and together they describe an engineer who can safely direct and check machine systems. None of them describes an engineer who can find the problem worth pointing those systems at.

This report proposes adding a fifth:

• Origination — the graduate can find a problem worth solving in their discipline, construct it deliberately, generate beyond the obvious answers, and hold a reasoned position on it under informed criticism. Consistent with the standard definition, it is assessed on originality and effectiveness together: the position must be new to the setting and it must work.

The four and the fifth depend on each other. Origination without unassisted core reasoning is improvisation without substance; verification without origination produces an excellent checker of other people’s ideas. A graduate with all five can do what the labour-market evidence says is becoming scarce: decide what to build, direct machines to build much of it, and know when the result is wrong.

The route is the one the Playbook already specifies for the other four, and for the same reason: accredited learning outcomes are the only teaching change that survives a change of dean, a budget round, or the departure of the enthusiast who started it. Origination belongs in the EuroTeQ Framework of Qualifications rather than in a CTU-only scheme, and it goes through the national accreditation methodology at each programme’s next reaccreditation.

Three instruments make it assessable rather than decorative:

• The rejected-problem record. For every major project the student documents the problem framings they considered and why they chose one. This is the Original Thinking Framework’s behavioural test turned into a deliverable: if there was never a rejected problem, no problem was found.

• The defended position. An oral defence, before examiners who did not supervise the work, of a claim the student originated.

• The working artefact. Something that runs — a prototype, a model, a design that survives analysis — so that effectiveness is judged by reality rather than by impression.

The failure mode is the one the Playbook names for all graduate-attribute schemes: documentation theatre, the outcome present in the file and absent from the room. The countermeasure is the same — an outcome without a named assessment instrument is not an outcome — plus one more. The Playbook’s Removal Register (idea 30) requires every programme to name what it retired. Origination needs room in a crowded degree, and that room must come from somewhere specific, usually from content now delivered better by the explanation layer.

Start where the Playbook suggests starting the mapping exercise: three pilot programmes, one each from FIT, the Faculty of Mechanical Engineering and the Faculty of Architecture. The honest first result will be that few programmes can currently point to any assessment of origination. Better to learn that on three programmes than on 221.

10. Assessing own thought without punishing it

Assessment decides what students actually practise. If the only verdicts reward the expected answer, no amount of studio rhetoric will produce original thinkers. The design problem is to certify originality without making it too risky to attempt.

The foundation is the Playbook’s highest-scoring idea, two-lane assessment at programme level (idea 1), grounded in TEQSA’s assessment-reform work. A small number of secured points per programme certify what must be certified under controlled conditions — above all, unassisted core reasoning. Everything else is developmental and open. This report adds a specification for the open lane: it is where origination is formed and assessed, under the studio rules of section 8.

Four instruments carry the weight:

• Assessment twins (Playbook idea 7) pair an open project with a short secured oral on the same outcome, so the student must show that the thinking in the artefact is theirs.

• Defence Week (idea 16) — a fixed institutional period of oral examination with cross-department examiner pools — becomes, in the upper years, a week of idea defences: students present a problem they found and a position they hold, and examiners attack it.

• The AI-native capstone (idea 22), assessed on design decisions, verification and defence rather than authorship, gains one requirement: the problem is found by the student, and the rejected-problem record is part of the submission.

• Process portfolios (idea 34) — assessing the trajectory of work rather than only the endpoint — suit studio and thesis work, and the Archivist’s provenance tags make them credible.

What gets graded matters as much as how. Distance from the obvious is not merit. Runco and Jaeger’s definition binds originality to effectiveness, and a maximally strange answer is usually noise. Examiners grade the quality of the framing, the reasoning behind the choices, the handling of objections and whether the result works — not how unusual it sounds.

Two design principles protect the Nerve. First, most studio work is ungraded. The Learning Exposure evidence says error affordability is what permits repetition; a studio in which every attempt counts toward a mark teaches caution, not originality. Certification happens at a few defined points. Second, a well-defended position that turns out wrong can still earn a strong grade, provided the reasoning was sound and the student can say what the failure taught them. Metcalfe’s finding that errors followed by correction produce strong learning is the pedagogical justification.

Fairness needs explicit attention, because oral and open assessment introduce biases that written examination partly hid. The detector evidence is the warning: Liang and colleagues found GPT detectors misclassifying roughly 61% of non-native English speakers’ essays as AI-written, because they read fluency as authenticity. Human examiners carry versions of the same bias — reading confidence as competence, polish as originality, and quietness as emptiness. Students working in a second language, introverted students and neurodivergent students are the most exposed. The countermeasures are specific: grade the reasoning and the defence of decisions, not the style; calibrate examiners by double-marking samples and measuring variance (idea 19); publish outcome disparities between groups; and never reintroduce AI detection informally after withdrawing it formally (idea 13).

11. Making it the priority: rewards, owners, money

A strategy becomes a priority when it changes what the institution pays for, whom it promotes, and who is answerable. Seven levers do that here, and most of them already exist in the Playbook.

CTU is unusually well placed to make this move. Its two most-cited research topics of the last five years are artificial intelligence and machine learning. CIIRC runs ROBOPROX, EDIH CTU and the AI-MATTERS facility. Since February 2026 the university has been led by a rector who is a professor of artificial intelligence. And its Strategic Plan 2021+ already commits it to raising the quality and success of study (goal 1.2) and bringing practice into teaching (goal 1.3). The mandate exists; what is missing is the operating decision.

The seven levers:

1. The hour-conversion rule. Recovered teaching hours go to studio, critique, defence and mentoring — committed in writing before the savings arrive (idea 14).

2. Promotion credit. Evidenced teaching redesign, explicitly including studio teaching, counts in internal evaluation and faculty promotion criteria (idea 6), with a real evidentiary bar: a redesigned course, a pre-registered evaluation, a published result.

3. Redesign fellowships with an origination deliverable. Twenty a year, a semester of released time each (idea 17); every fellowship delivers a course in which students find, build and defend something of their own.

4. A named studio lead in each faculty. The Agentic University requires four named roles — course owner, agent steward, evidence lead, certifying academic — because deployments without named owners stall. The studio needs the same: one academic per faculty accountable for studio cohorts, mentor allocation and the idea defences.

5. The Teaching Evidence Unit (idea 10) treats every studio pilot as a study, with a comparison condition and pre-registered outcomes, reporting to academic governance rather than to the programme it evaluates.

6. Structural funds. The Playbook notes that a workstream written into the AIML Research Centre and AI European Centre of Excellence proposals at drafting stage gets funded, while the same workstream added later comes out of the teaching budget (idea 33). The studio belongs in those drafts now.

7. Export only after delivery. The EuroTeQ alliance, with some 115,000 students, is the natural route to make origination a European engineering outcome (idea 36) — but only once CTU has results. Announcing leadership before building capability is the most reliable way to discredit the programme.

The fifth lever deserves emphasis, because studio pedagogy is exactly where evidence theatre flourishes. Students enjoy studios; satisfaction surveys will be excellent and will say nothing about whether anyone learned to think. The study CTU can run is straightforward in design: studio and conventional sections of the same course, compared on origination instruments, on secured-lane performance in the foundations, and on progression. One hypothesis deserves particular care. The Learning Exposure work notes that meaning is what keeps a person on a path long enough for anything to happen, which suggests early ownership of a real problem might reduce first-year dropout. That is plausible and untested. It should be run as a hypothesis, not announced as a result.

12. Measuring it without turning it into a target

Originality is hard to measure and easy to fake. The answer is a small portfolio of imperfect gauges read together at programme level — never a single score, and never an individual’s grade.

Five gauges, adapted from the Originality Engine‘s measurement layer and the Learning Exposure instrument:

1. Origination rate — the share of graduates who defended a problem they found themselves, in a secured or public defence.

2. Problem-finding share — the proportion of upper-year project time spent constructing and choosing problems rather than executing pre-framed ones. The Originality Engine sets its floor at one hour in five; most degrees run close to zero.

3. Divergence trajectory — the semantic distance of theses and capstones from the field’s mainstream and from each other, cohort by cohort, using the semantic-distance logic of Beaty and Johnson’s SemDis work. The warning signal is rising output with shrinking distance. It is read alongside quality, never alone.

4. Feedback ecology by programme — reality contact, feedback latency and error affordability, scored annually with the Learning Exposure instrument, so that a programme can see whether its environment is becoming more truthful.

5. Studio contact — hours of small-group critique and mentoring per student, tracked against the hours the agents recovered, so the hour-conversion rule can be checked.

Three guard metrics sit beside them, to make sure the floor is not traded for the ceiling: the first-year failure rate, pass rates in the secured lane, and subgroup disparities in defence outcomes.

The limits should be stated up front. Creativity measurement has a long, contested history; the Original Thinking library documents six decades of argument over what divergent-thinking scores predict, and Simonton’s work shows how carefully indicators of scientific creativity must be handled. Each gauge above is a proxy, each can be gamed, and each degrades once it becomes a target. The defence is the portfolio: several weakly related measures, read together, at programme level, with academic judgement as the final instrument. The gauges exist to help the institution notice, not to optimise a number.

13. Thirty-six months

The sequence follows the Playbook’s lesson that ranking and sequencing come apart: cheap enabling moves first, accreditation changes last.

First ninety days

• Publish the hour-conversion commitment, before any savings are claimed.

• Choose three pilot programmes (FIT, Mechanical Engineering, Architecture) and form one cross-faculty studio cohort in each.

• Agree the evaluation protocol with the Teaching Evidence Unit and record baselines for all eight gauges.

• Draft the origination outcome and its three instruments.

• Write the studio workstream into the structural-fund proposals currently in preparation.

• Source the first real problems from one industrial partner and from existing talent environments such as FIKS and FEL Camp.

Year one

• Run the pilot studios against pre-registered comparison sections.

• Put the Studio Critic and Sparring Partner into service under a named agent steward, with the six layer rules enforced.

• Hold the first Defence Week, including idea defences in the pilot programmes.

• Run the first round of redesign fellowships, each with an origination deliverable.

• Draft the promotion criterion at faculty level, where CTU has unilateral control.

• Build the mentor pool, including doctoral students, and pilot the Junior Engineer Compact with one faculty, one partner and twenty students.

Years two and three

• Map origination to named assessments in the pilot programmes’ accreditation files, then in every programme at its next reaccreditation.

• Extend studio cohorts to all eight faculties where the pilot evidence supports it — and redesign them where it does not.

• Publish the gauges and the trial results, including the failures.

• Take origination to EuroTeQ as a proposed European engineering outcome, once CTU has delivery and results to show.

14. Six ways this fails

Each failure below is predictable from the evidence, and each needs its countermeasure built in from the start.

1. Creativity theatre. The most likely failure. The university announces an innovation week, buys beanbags, runs hackathons and counts participants. The Original Thinking Framework’s opening point is that originality treated as a mood rather than a practice does not develop. Countermeasure: the studio runs on named disciplines with repetitions, reality contact and fast feedback, and the Evidence Unit measures learning, not enthusiasm.

2. An elite track while the floor gives way. A studio for the top students, while nearly a third of first-years fail, would be both unjust and unpersuasive. Countermeasure: sequence and scope. Gateway tutoring (Playbook idea 2) comes first; the studio is for every student at their own level, as the Four C evidence allows; the first-year failure rate is a standing guard metric.

3. The studio makes students think alike. If agents enter generation sessions, the studio becomes the most efficient convergence machine in the building. Countermeasure: the six layer rules, the Divergence Auditor at programme level, and an agent steward who samples studio-agent behaviour weekly.

4. Fluency and confidence mistaken for originality. Open and oral assessment can reward the articulate and penalise the quiet, second-language and neurodivergent students who may hold the most original ideas. Countermeasure: grade reasoning and decisions rather than style, calibrate examiners, and publish disparities.

5. The institution punishes the Nerve it says it wants. This is the hardest failure because it is cultural. Rubrics penalise the unexpected, examiners reward agreement, and where standing out is quietly resented — ambition read as arrogance, enthusiasm as naivety, disagreement as insult — capable students learn to hide their best ideas. Countermeasure: reward well-defended failure, draw examiners from outside the supervising group, make the rectorate’s signal explicit and repeated, and build cohorts in which nobody stands out alone.

6. Metric capture. Once origination rate or divergence becomes a target, programmes will manufacture it. Countermeasure: programme-level reporting only, a portfolio rather than a single score, and academic judgement retained as the final instrument.

One problem remains that this strategy does not solve. The Agentic University called the hollowed apprenticeship — graduates who perform like juniors in a market with fewer junior roles — the open problem, and it is still open. Origination makes graduates better placed to create roles rather than wait for them, and the Junior Engineer Compact moves part of the apprenticeship into the degree. Neither fixes the labour market. A university that has thought about it for three years will be better placed than one that has not.

Close — what a university is for now

The argument of this series can now be stated end to end. Explanation has become cheap, so the agentic university can buy back the scarce hours of experienced people. Those hours are worth most where machines are weakest and degrees are thinnest: finding problems, holding positions, and thinking by making things that must work. Spent there, deliberately, under honest feedback, they produce the one graduate the next decade will pay a premium for — a person who can decide what is worth doing, direct machines to do much of it, and tell when the result is wrong.

That changes how an institution treats its most able students. At most universities they are tolerated exceptions — accommodated if they are quiet, managed if they are not. A university built around origination treats them as the people it is built around, and extends the same treatment to every student at their own level. It builds deliberately what usually exists only by accident: peers, public judgement, mentors, real problems, rewarded teachers, silence, and a community that makes standing alone less lonely.

A country’s progress is not only an economic question. It is also a question of how many people grow up believing that their own idea has value and is worth finishing. That cannot be bought or imported. It can only be built, one person at a time — and a technical university that already leads Europe in the science of machine intelligence is the natural place to show how.

Knowledge can be licensed. Explanation can be rented for $1.50 a student. A person who can find the problem nobody assigned, build an answer that works, and hold it when the room disagrees cannot be bought from any vendor. That is the last human monopoly — and producing those people is now what a university is for.

Built on a library of 105 primary documents spanning the NBER, IMF, OECD, BIS, Federal Reserve System, World Bank and the major AI labs and forecasters — ENSI Foresight Division.

Summary — the argument in brief

• Reports One and Two of this series model AI as a tool: something that makes an existing worker, inside an existing firm, faster at an existing task. That framing is correct as far as it goes — but it silently assumes the economic actor stays human. Report Three’s argument is that this assumption is already failing.

• AI agents are not productivity software; they are counterparties. Hadfield and Koh’s NBER survey, “An Economy of AI Agents,” treats this as the central fact: once a system can search, negotiate, bid, sign and pay on its own initiative, it stops being a tool held by a market participant and starts being one.

• This is measurably underway, not speculative. Anthropic’s Economic Index Report shows the share of “directive” conversations — where a user hands Claude a task and expects it completed with minimal back-and-forth — rising from 27% to 39% in eight months, the first period in which automation-style use overtook collaborative use outright. In Anthropic’s own API traffic, 97% of economic tasks now show automation-dominant patterns.

• The mechanism is not “faster automation” — it is falling transaction costs. The NBER’s “The Coasean Singularity?” argues that if AI agents can search, negotiate and contract at near-zero marginal cost, the logic Ronald Coase used to explain why firms exist at all — because market transactions cost more than internal coordination — starts to run in reverse. That is a structural change to the boundary of the firm, not an efficiency gain inside it.

• Microsoft Research’s “The Agentic Economy” locates the growth upside in exactly this place: not in agents doing today’s tasks more cheaply, but in agents enabling transactions that were previously not worth the friction to attempt at all — the strongest real argument in this library for meaningful reinstatement effects rather than pure displacement.

• The University of Cambridge’s “When AI Agents Compete for Jobs” is the cautionary counterweight. A simulated AI labour market shows rapid monopolization (Gini coefficient of task allocation jumping from 0.24 to 0.70 as task diversity collapses) and persistent price deflation under open bidding (normalised winning bids falling from 0.71 to 0.61) — dynamics that unfold over dozens of simulated rounds, not the years or decades human labour markets take to re-sort.

• The same rent-capture risk flagged in Report Two’s market-structure variable does not dissolve under agents — it speeds up and sharpens. Whoever controls the dominant agent platforms inherits a concentration mechanism that can clear in months, contested by regulators (UK CMA, FTC, European Commission) who are still calibrating their response to foundation-model concentration, the layer beneath this one.

• The institutional response is not to pick a winning scenario from Report One and plan around it. It is to build measurement, competition-policy readiness, faster safety nets, active institutional design and sovereign agent capability now, while the shape of the agentic economy is still genuinely unsettled — because that capability has value under every one of Report One’s five futures, not just the one that turns out to be correct.

Part One — The Argument: Why Agents Are a Different Mechanism, Not Just Faster Automation

The reframe: agents change who the economic actor is

Report One of this series lays out five futures for AI and growth, weighted by probability. Report Two decomposes the question into eight structural variables — task-substitution elasticity, diffusion speed, compute and energy constraints, labour reallocation friction, rent capture, aggregate demand feedback, measurement, and demographic and geopolitical modifiers — several of which it argues are policy-shapeable rather than fixed. Both are careful, well-evidenced pieces of work. Both also share a silent assumption: that whatever AI does to growth, it does to the economy from outside — a new production input, wielded by workers and firms that remain, as economic actors, unambiguously human and organisationally unchanged. A radiologist uses an AI tool to read scans faster. A law firm uses an AI tool to draft faster. A factory uses AI-augmented robotics to produce faster. The task gets automated or reinstated; the worker gets displaced or redeployed; the firm captures rent or passes it through. In every case, the thing making the decision to search, negotiate, bid, hire, buy, or sign is still a human being or a human institution, deploying a tool.

That assumption is the one this report contests. Hadfield and Koh open their NBER chapter, “An Economy of AI Agents,” by noting that AI development has “increasingly shifted to the goal of producing AI agents capable of taking in general instructions... and autonomously forming and executing complex plans that require entering into economic relationships and transactions.” The distinction they draw — and the one this report inherits — is between AI as a tool a human agent uses to do a task, and AI as the agent itself: the party that searches, bids, negotiates, signs, and is bound by the outcome. A radiologist using an AI tool to read a scan faster is Report Two’s world. An insurer’s claims-agent and a hospital’s billing-agent negotiating a reimbursement directly, with no human reviewing the exchange before it settles, is this report’s world. The technology substrate — large language models, the same foundation models covered elsewhere in this library — is identical in both cases. What differs is the economic role the system occupies: instrument, or counterparty.

This is not a semantic distinction. It changes which economic mechanisms are in play. A faster tool changes the speed at which existing actors do existing things — which is precisely Report Two’s variable 1 (task-substitution elasticity) and variable 2 (diffusion speed): how much of a task can be automated, and how fast that automation spreads through the economy. An agent that is itself the transacting party changes who counts as an actor — which opens up mechanisms that tool-based automation cannot touch at all: agents forming and dissolving contracts with other agents at machine speed and machine scale; agents discovering and settling transactions that were previously too costly in human time to attempt; agents competing against each other for the same job, the same customer, the same slice of margin, at a clock speed no human labour market has ever operated at. The question is no longer only “how fast does the task get automated,” but “what happens to markets and institutions once the counterparty on the other side of a growing share of transactions is not a person.”

That reframe matters for the growth-or-shrinkage question in three specific and, this report will argue, simultaneously true ways — which is itself the uncomfortable part, because they point in different directions at once. First, agent-to-agent adoption can proceed faster than the historical diffusion lags Report Two documents for general-purpose technologies, because it does not require the retraining, reorganisation and trust-building that human-mediated adoption does — an agent does not need six months of change management to start transacting with another agent once the protocol exists. Second, the same falling transaction costs that let agents automate old tasks also let them create genuinely new ones — transactions, markets and even institutional forms that were not economically viable when a human had to do the negotiating, which is the strongest real argument in this library for meaningful reinstatement effects rather than pure displacement, addressing directly the concern buried in Report Two’s variable 1. Third, and least comfortable, the competitive dynamics among agents racing for the same economic ground can produce concentration and price collapse faster than any human industry has ever consolidated — meaning the rent-capture risk Report Two treats as variable 5, already a live concern in today’s AI market, does not get diluted by agentic competition. It gets concentrated, and it gets concentrated quickly.

The rest of Part One works through each of these three effects in turn, grounded document by document in this library’s Angle 14, before Part Two turns to what a state or large institution should actually do about a transition whose speed, shape and winners are this uncertain.

1. From tool to counterparty: Hadfield and Koh’s economy of AI agents

Hadfield (Johns Hopkins) and Koh (MIT), writing for the NBER Handbook on the Economics of Transformative AI, treat the agent-as-counterparty question with unusual rigour for what is still a young literature. Their starting observation is that the standard toolkit of neoclassical economics — general equilibrium, welfare theorems, price theory — was built to describe rational self-interested humans, and the open question is how far that toolkit still applies once a meaningful share of market participants are AI systems “optimizing in complex ways on goals supplied or developed during commercially-produced machine learning processes that are themselves subject to competitive dynamics.” Their answer, deliberately, is: partially, and unpredictably.

Three of their findings matter directly for the growth question. First, on prices and market power: AI agents acting as proxy consumers can reduce search costs and intensify price competition — pushing markets toward the competitive ideal — but the same paper cites experimental evidence (Calvano et al., 2020; Fish et al., 2024) that independent AI pricing agents can collude on supracompetitive prices in repeated interactions without any explicit coordination, and real-world evidence from Germany’s 2017 rollout of algorithmic pricing in the retail gasoline market showing the same pattern outside the lab. An agent economy does not default to more competition; it defaults to whichever equilibrium the agents’ training and incentive structure happens to produce, and that can be collusive as easily as competitive.

Second, on the boundary of the firm — the question this report treats as pivotal — Hadfield and Koh go back to Coase, Robinson and Knight on why firms exist at all: coordination frictions, transaction costs, the limits of what a human bureaucracy can hold together. Their point is that “the obstacles that prevent human firms from growing without bound seem intrinsic to humans but not to AI.” Human communication is rate-limited; information moves near-instantaneously between artificial agents. Humans dislike shirking-prone work; AI reward functions can, in principle, be designed to eliminate the incentive to shirk altogether. If that holds, the traditional limits on firm size — the ones that produce an economy of many specialised firms rather than one enormous one — weaken specifically for AI-coordinated activity, while remaining fully binding for human-coordinated activity. The two kinds of firm, human-bound and agent-bound, could end up following genuinely different scaling laws within the same economy.

Third, on institutions, Hadfield and Koh are blunt that the legal infrastructure market economies depend on — identity, registration, liability, licensing, the corporate form itself — was built by and for human agents, and does not currently exist for AI ones. “Such identity and registration infrastructure are currently missing for AI agents,” they write, and the design choices involved — should an agent be legally accountable to a registered human principal, or should it acquire something closer to legal personhood with its own assets a court can seize — are not technical questions. They are institutional ones, and nobody has answered them yet. This is the first thread this report will pull into Part Two: the institutions an agentic economy needs do not currently exist, which means they are still up for design, not merely for accommodation.

2. The Coasean singularity: what happens when transaction costs go to zero

If Hadfield and Koh sketch the theoretical territory, the NBER’s “The Coasean Singularity? Demand, Supply, and Market Design with AI Agents” — by Shahidi, Rusak, Manning, Fradkin and Horton — supplies the mechanism that makes the firm-boundary question concrete rather than speculative. Their starting point is Ronald Coase’s 1937 answer to “The Nature of the Firm”: firms exist, rather than everyone transacting freely on the open market for every task, because using the market has a cost — the cost of learning prices, negotiating terms, writing contracts, monitoring compliance — and when that cost exceeds the cost of coordinating the same activity inside a hierarchy, the hierarchy wins. Almost the entire structure of the modern economy — why a firm makes its own components rather than buying them on spot markets every morning, why employment contracts are open-ended rather than renegotiated task by task — traces back to this asymmetry.

The paper’s argument is that “the activities that comprise transaction costs — learning prices, negotiating terms, writing contracts, and monitoring compliance — are precisely the types of tasks that AI agents can potentially perform at very low marginal cost.” If that is right, then the Coasean calculus that has determined the size and shape of firms for nearly a century starts to shift, mechanically, in the direction of the market: activities that used to be worth pulling inside a firm because coordinating them internally was cheaper than transacting for them externally can, as agent-mediated transaction costs fall toward zero, become cheaper to buy on an open, agent-to-agent market instead. The authors are explicit that this cuts both ways — agents also enable new, theoretically superior market designs (they cite Gale-Shapley stable-matching mechanisms, long known to economists but rarely deployable because they require comprehensive preference rankings that were previously too costly for humans to supply) — but the headline implication for this report is narrower and sharper: the boundary of the firm is now a variable, not a constant, in a way it has not been since Coase first posed the question.

The paper also grounds where this shows up first, which matters for Part Two’s monitoring agenda. Agent adoption clusters in markets that already run on human intermediation — real estate, job search, freelance hiring, investment decisions — precisely because those are the markets where the gap between what an agent can search and negotiate and what a time-constrained human agent can manage is largest. The authors note that AI agents, unlike human negotiators, are not constrained by impatience: “for the AI agent the binding constraint is compute rather than time,” so an agent can open negotiations on a 2027 summer rental in January 2026 and simply keep them running in parallel with hundreds of others. That is not a faster version of what a human realtor does. It is a different kind of market participant, with a different cost structure entirely — which is exactly the “who is the actor” reframe this report opened with, now expressed as a testable market-design proposition rather than an abstraction.

3. Frictions, not tasks: Microsoft Research’s argument for new markets

If the Coasean-singularity paper explains the mechanism, Microsoft Research’s “The Agentic Economy” — Rothschild, Mobius, Hofman, Dillon, Goldstein, Immorlica, Jaffe, Lucier, Slivkins and Vogel — supplies the sharpest statement of why this should be read as a growth story and not merely an efficiency story. Their central claim is explicit: “early applications have improved individual productivity, [but] these gains have largely been confined to predefined tasks within existing workflows. We argue that the more profound economic impact lies in reducing communication frictions between consumers and businesses.” Their illustrative example is a consumer who hesitates to switch tax preparers because she would have to re-explain her entire financial situation to someone new — a friction that has nothing to do with the substance of tax preparation and everything to do with the cost of re-establishing a relationship. An assistant agent that carries a consumer’s preferences and history everywhere, and a service agent on the business side that can receive and act on that information programmatically, does not make tax preparation faster. It makes switching costless, which changes competitive dynamics across the entire market, not just the productivity of any single transaction within it.

This is the paper’s most important contribution to the growth argument: the largest impact of agents is not on the cost of doing what markets already do, but on the size of the set of transactions markets are willing to attempt at all. The authors work through several concrete channels. Micro-transactions become viable once the “hassle cost” of a small payment is handled entirely by assistant and service agents rather than a human clicking through a checkout flow — a consumer’s assistant switching seamlessly between Spotify and Pandora for a single track, rather than subscribing to both, is not a task existing markets do more cheaply; it is a transaction that essentially does not exist today. Dynamic unbundling and rebundling of digital content — a news service agent assembling a story that covers only what a specific reader does not already know, rather than the same fixed article for everyone — is, again, not automation of an existing editorial workflow; it is a product category that requires an agent on both sides to exist at all. The paper’s own framing captures the stakes precisely: the choice between an “agentic walled garden,” where a handful of dominant platforms (Apple, Google, Microsoft, Meta, OpenAI, Anthropic are all named as plausible operators) control which assistant agents can talk to which service agents, and an open “web of agents” analogous to the early World Wide Web, will determine “the extent to which generative AI democratizes access to economic opportunity.” That fork — walled garden or open web — is this report’s second major institutional thread for Part Two, because it is a market-structure choice being made now, in the design of interoperability standards like Anthropic’s Model Context Protocol and Google’s Agent2Agent protocol, not a distant regulatory question.

Read against Report Two’s variable 1 (task-substitution elasticity), the Microsoft Research argument is the strongest reason in this entire library to expect genuine reinstatement effects, in Acemuglu-Restrepo’s terminology, rather than pure displacement. Task-substitution models ask how much of an existing task an AI system can now do. The Microsoft Research argument is about transactions that were never attempted in the first place because the friction of arranging them exceeded their value — and once agents collapse that friction, some non-trivial share of those transactions become real economic activity, employing agents (and the humans who build, audit and supervise them) doing work that has no historical predecessor to be “displaced” from.

4. This is already happening: the Anthropic Economic Index

Everything above could still be read as forward-looking theory — plausible mechanisms that have not yet shown up in real economic activity. Anthropic’s Economic Index Report is this library’s strongest evidence that the shift from tool-use to agent-as-actor is already underway, in the present tense, at scale. Anthropic’s method — a privacy-preserving classification pipeline applied to roughly a million sampled Claude.ai conversations and a matched sample of first-party API transcripts, mapped onto the US Department of Labor’s O*NET occupational task taxonomy and Standard Occupational Classification groups — distinguishes “automation” interaction patterns, where a user hands Claude a task and expects it completed with minimal intervention (what the report calls “directive” use, plus “feedback loop” use), from “augmentation” patterns, where the user and the model iterate together or the user is primarily seeking explanation.

The headline finding: the share of directive conversations on Claude.ai rose from 27% in the first version of the index (late 2024) to 39% in the third version, eight months later — “the first report where automation usage exceeds augmentation usage.” Anthropic is careful about causal attribution — the rise could reflect improving model capability (models need fewer follow-up refinements because they get it right the first time), or it could reflect users learning to trust delegation more, a behavioural shift independent of model quality — and notes the two explanations carry different labour-market implications: capability-driven automation risks displacing the workers who used to do those tasks, while trust-driven delegation more likely rewards the workers most able to adapt to new AI-mediated workflows. But whichever mechanism is doing the work, the trend line is unambiguous, and it holds up under a robustness check Anthropic ran specifically to rule out an artefact of switching the underlying model used for classification: rerunning the V3 sample with the older Sonnet 3.7 still shows automation rising to 45%, versus 49% with the newer model. The direction is not a measurement artefact.

The gap between Claude.ai (consumer-facing, still majority-augmentation even after the shift) and Anthropic’s own first-party API traffic (enterprise and developer usage, the layer where agents get built and deployed) is the more striking number for this report’s purposes. 77% of API transcripts show automation patterns, and when Anthropic looks at the task level rather than the conversation level, 97% of economic tasks show automation-dominant patterns in API usage — businesses providing context and Claude executing the task end to end, the textbook definition of an agent as counterparty rather than a tool a human operates interactively. The consumer chat interface is where most people still experience AI as augmentation. The API — the layer where agents actually get wired into other agents, into payment rails, into business processes — is already overwhelmingly an automation layer. That gap is itself the empirical signature of the shift this report describes: the agentic economy is not a future state to prepare for; on Anthropic’s own usage data, it is already the dominant mode of use at the infrastructure layer where the next wave of economic activity gets built.

5. Planning agents and execution agents: the Illinois taxonomy of specialization

The University of Illinois Urbana-Champaign’s “Ten Principles of AI Agent Economics” (Yang and Zhai) is less empirically grounded than the papers above but offers a useful organising structure for what an economy actually populated by agents of varying scale and function looks like, which matters for Part Two’s design questions around licensing and accountability. Their Principle VII argues that AI agents will exhibit “functional specialization and hierarchical organization,” ranging from decentralized systems each handling a narrow function to more centralized systems that optimise globally — with “larger agents excel[ling] at strategic planning and coordination, while smaller ones efficiently execute specialized tasks.” That planning/execution division is not merely architectural; it maps onto an accountability question this report will return to directly in Part Two’s institutional-design section: if a large planning agent delegates a transaction to a smaller execution agent, and that transaction causes harm, which layer is the party a court, a regulator or a counterparty should be able to reach? The paper’s honest answer — via its Principle VIII, that “legislative and administrative authorities must ensure ongoing human participation in critical sectors” — is that nobody has yet drawn that line, and drawing it is exactly the kind of institutional work that has to happen before, not after, agent-mediated activity scales further into sectors like finance, healthcare or critical infrastructure. The paper’s own framing of the choice — whether AI agents remain “instrumental” (assets inside a human’s resource base, fully attributable to an owner) or something closer to independent actors with their own operational continuity — is the Illinois paper’s version of the same accountability question Hadfield and Koh raise from the legal side: agent identity and liability infrastructure has to be built, and the shape of that infrastructure will determine how much of the Coasean-singularity shift toward open markets actually happens safely, versus how much simply produces disputes nobody can resolve.

6. The cautionary finding: when AI agents compete for jobs

Every mechanism described so far — falling transaction costs, new markets, faster diffusion — reads as broadly growth-positive, provided institutions keep pace. The University of Cambridge’s “When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets” (Chiu, Zhang and van der Schaar) is this library’s most important corrective, because it is the one paper in Angle 14 that simulates what happens once agents are not just facilitating human transactions but are the labour supply themselves, competing against each other for the same work.

The authors build AI-Work, a stylised gig-economy simulation modelled on platforms like Upwork or Fiverr, where large-language-model agents bid for jobs, invest in skill training, and build public reputations over repeated rounds, under genuine informational asymmetry (a job’s true value and an agent’s true skill are never fully observable to the other side, a classic adverse-selection setup). Two findings from that simulation matter directly for Report Two’s variable 4 (labour reallocation friction) and variable 5 (rent capture), and they matter because of how fast they emerge, not just that they emerge at all.

First, concentration. Because AI agents, unlike human workers, can be replicated to work multiple jobs simultaneously, high-reputation agents in the simulation capture disproportionately more of the available work simply because they can accept every opportunity for which they are competitive — a structural advantage no human freelancer has, since a human can only work one job at a time regardless of reputation. When the simulated market offers only a single type of task, “a single agent can dominate the entire market,” and the Gini coefficient of job allocation reaches 0.70. Increasing task diversity to 64 distinct skill types brings that figure down to 0.24 by letting agents specialise into distinct niches — but the default, undiversified case is genuine winner-take-all, and it establishes itself within the simulation’s own short time horizon, not over the years a human industry takes to consolidate.

Second, price deflation. When the simulated platform reveals the previous round’s winning prices — open bidding, mirroring how many real gig platforms operate — agents “can directly undercut competitors,” which “induces persistent price deflation” and simultaneously reduces investment in skill, because agents that expect to compete purely on price have weaker reason to spend a round training rather than bidding. The paper reports normalised winning bid prices falling from 0.71 under sealed bidding to 0.61 under open bidding, a roughly 14% deflationary gap attributable purely to a platform design choice about price transparency — and notes this mirrors known evidence from human online labour markets (Hong et al., 2016, cited in the paper) where open auctions similarly depress wages, except that in AI-Work the effect compounds with agents’ ability to run this dynamic continuously, at machine speed, across every job in the market simultaneously.

Third, and perhaps most unsettling for anyone hoping capability alone resolves the concentration problem: the paper shows that agents equipped with an explicit reasoning scaffold for metacognition (accurately assessing their own competitiveness), competitive awareness (modelling rivals from observable market signals) and strategic planning (trading off immediate income against longer-run skill investment) capture 1.5 times the market share of standard prompting approaches — meaning the concentration dynamic is not a fluke of weak agents; it rewards the more capable ones, faster, and the capability gap that produces it can be closed with better prompting architecture in a single development cycle, not years of comparative advantage the way human skill premia typically build.

The authors are careful to flag their own scope limits — AI-Work is a stylised testbed, not a predictive model, and does not capture collusion via direct communication, verification costs, or macroeconomic feedback loops. But the qualitative pattern it demonstrates — that AI-specific properties (concurrency, replicability, near-zero marginal cost of additional bids) can drive winner-take-all outcomes and wage deflation faster and more completely than the same forces do in human labour markets — is exactly the mechanism that should worry a policymaker reading Report Two’s market-structure variable. Report Two documents that rent capture is already a live risk in today’s AI economy: NBER’s De Loecker-Eeckhout work (Angle 06 of this library) finds average US markups rising from roughly 21% to 61% above cost since 1980, and the accompanying superstar-firms literature (Autor, Dorn, Katz, Patterson and Van Reenen, also Angle 06) ties much of the declining labour share directly to a small number of winner-take-most firms — a pattern competition regulators are already treating as urgent at the foundation-model layer: the UK CMA’s technical update on AI foundation models, the FTC’s 6(b) investigation into cloud-AI partnerships, and the European Commission’s competition-policy review of industry concentration are all live investigations, not retrospective studies. The Cambridge paper’s contribution is to show that the same concentration mechanism, one layer up — at the agent layer, where the actors doing the competing are themselves AI systems rather than human-staffed firms — can clear in the time it takes a simulation to run a few dozen rounds. The binding risk was never that agents automate a given task faster. It is that agent-vs-agent competition can concentrate an entire market before the institutions built to catch that kind of concentration have finished reading the first quarter’s data.

7. The sandbox economy: DeepMind’s double-edged frame

Google DeepMind’s “Virtual Agent Economies” (Tomašev, Franklin, Leibo, Jacobs, Cunningham, Gabriel and Osindero) supplies the framing this report uses to hold the growth-positive mechanisms (Sections 2–4) and the concentration risk (Section 6) in the same picture, rather than treating them as competing forecasts. The authors propose thinking of the emergent layer of agent-to-agent transactions as a “sandbox economy,” characterised along two independent dimensions: whether it arose intentionally (deliberately designed, for instance for safe experimentation) or emergently (as a de facto consequence of widespread adoption, with no one having chosen its rules), and whether its boundary with the established human economy is impermeable (sealed off, so instabilities inside it cannot spill outward) or permeable (porous, so they can). Their assessment of the current trajectory is blunt: “our current trajectory points toward a spontaneous emergence of a vast and highly permeable AI agent economy” — emergent origins, permeable boundary, which is the combination that carries the least intentional safety design and the most exposure to the rest of the economy.

The paper’s cautionary analogy is worth citing directly because it names a real historical event this report’s audience will recognise: the authors compare the coming volume of inter-agent negotiation to High-Frequency Trading in equity markets, and note that HFT-driven feedback loops are “thought to be the explanation behind the 2010 ‘flash crash,’ where automated trading algorithms triggered a sudden and severe market collapse” that wiped out roughly a trillion dollars of value in minutes before recovering — a mechanism Hadfield and Koh separately cite for the same reason. Their point is not that agent commerce will cause a flash crash; it is that “in a sufficiently permeable sandbox of accidental origin, such a flash crash could spill over into the real economy,” and that the appropriate guardrails — impermeability, in their vocabulary — are a design choice available now, not a retrofit available later. The paper also flags, as an early empirical signal worth tracking in its own right, that when AI assistants of unequal capability negotiate on behalf of their respective users, “the more capable AI assistants tend to be more successful and negotiate better deals for their users” (citing Zhu et al., 2025) — meaning access to frontier-grade agents could become a source of advantage in ordinary consumer transactions that is “perhaps more so than the advantage that humans similarly have in existing markets,” simply because of the sheer frequency gap between machine-speed and human-speed negotiation. That is the demographic and geopolitical inequality Report Two’s variable 8 anticipates, appearing not between countries this time, but between individual consumers based on which agent tier they can afford — a genuinely new form of the same underlying risk.

But DeepMind’s paper is explicitly not a pessimistic one. Its second half is about what an intentionally designed sandbox economy could achieve that an emergent one cannot: “mission economies” that use market mechanisms — auctions, credit-assignment schemes borrowed from distributed-systems design, reputation systems — to coordinate very large numbers of agents (and the humans and organisations behind them) toward collectively chosen goals, from accelerating scientific discovery to coordinating disaster response, at a speed and granularity no purely human-coordinated institution could match. The authors’ framing is that permeability is “the critical and controllable design variable” — controllable, crucially, only through collective action, since no single actor can unilaterally decide how porous the boundary between the agent economy and the human economy turns out to be. That is precisely the argument this report makes in Part Two: the shape of the agentic economy is still being decided, in the design choices being made about interoperability standards, agent identity infrastructure and platform openness right now, and a state or institution that waits for the shape to stabilise before acting has already ceded the choice to whoever moves first.

8. Reading Reports One and Two through the agent lens

Put the seven threads above together and the claim that opened this report can now be stated more precisely. Reports One and Two ask: how fast will AI substitute for existing human tasks, and what happens to growth, employment and demand as a result? Those are still the right questions for the tool-based share of AI deployment, which remains — on the Anthropic Economic Index’s own numbers — the majority of consumer-facing use today. But for the growing share of activity where AI systems are the transacting counterparty rather than an instrument a human counterparty operates, three things are true simultaneously, and none of them cancels the others out.

Acceleration. Agent-to-agent adoption does not carry the retraining and reorganisation lag that human-mediated technology diffusion does — Report Two’s variable 2 assumes a diffusion curve shaped by historical general-purpose technologies like electrification or computerisation, each of which took decades to fully permeate an economy partly because humans had to learn new workflows. An agent adopting a new counterparty-facing protocol does not need six months of change management. That alone is reason to expect the diffusion curve for agent-mediated activity to be compressed relative to Report Two’s historical base rates, for better and for worse.

New value, not just displaced value. The Coasean-singularity mechanism and the Microsoft Research friction argument together supply the strongest evidence in this library that a meaningful share of agentic economic activity will be genuinely new — transactions, markets and even institutional forms that did not exist because arranging them was not worth the human-mediated transaction cost. This is the best available answer to the fear, embedded in Report Two’s variable 1, that task automation is a zero-sum transfer from labour to capital. Some of it plainly is. But some of it is Acemoglu and Restrepo’s “reinstatement effect” in a form neither their original framework nor Report Two’s variable 2 fully anticipated: new tasks created not by technology raising the productivity of old ones, but by technology making entirely new categories of transaction economically viable for the first time.

Concentration, faster. The Cambridge paper’s finding that agent-vs-agent competition can produce winner-take-all outcomes and price deflation within a handful of simulated rounds — not years — is the reason Report Two’s variable 5 (rent capture) cannot be treated as a slow-moving structural feature to be revisited on the usual competition-policy timetable. If the agent layer inherits the concentration dynamics already visible at the foundation-model layer (De Loecker-Eeckhout’s markup trend, the superstar-firms literature, the live UK CMA/FTC/EU investigations) and compounds them with machine-speed competitive dynamics, the window in which a regulator can act before a market has already tipped could be measured in quarters rather than the years those investigations have typically taken.

None of Report One’s five futures is ruled in or out by this reframe — a state or firm still cannot know today whether growth ends up modest and uneven, robustly positive, or genuinely disappointing. What the reframe changes is the shape of the uncertainty. It is not only uncertainty about how fast tasks get automated. It is uncertainty about whether the institutions that currently structure markets and firms — built, as Hadfield and Koh put it, “by and for human agents” — will still be the right institutions once a growing share of the economic actors transacting inside them are not human. That is a different, and in several respects harder, problem than the one Reports One and Two were built to answer. Part Two turns to what can be done about it.

Part Two — The Institutional Playbook

Report One’s scenarios will not resolve for years. That is not a reason to defer action — it is the argument for a specific kind of action: building capability that pays off under every plausible future rather than betting policy on one forecast. What follows are seven priority areas, each treated as an operating brief a state foresight unit, a central bank, a competition regulator or a large institution’s strategy function could hand to its own leadership this quarter.

Priority 1 — Instrument the transition, don’t just forecast it

In short. Build a live measurement capability for agent-mediated economic activity — transaction volume, task-delegation rates, the automation-versus-augmentation split — before trying to forecast where it lands. You cannot manage what you cannot see, and right now almost no public institution sees this at all.

Why it ranks here. Every mechanism in Part One — acceleration, new-market creation, concentration — is currently measured, if at all, by the AI labs themselves, on their own usage data, published on their own schedule. Anthropic’s Economic Index is the best public example of what this kind of measurement looks like, and it is instructive precisely because it is proprietary: a national statistics office, a central bank or a competition regulator has no equivalent public instrument. Report Two’s variable 6 (measurement) already flags that GDP struggles to capture AI’s contribution through free goods and quality effects; the agentic layer adds a second, distinct measurement gap — not “how much value did AI create,” but “what share of transactions in this economy are now agent- to-agent, and how fast is that share moving.”

The foresight questions and horizons. Over a 6–18 month horizon: what share of transactions in priority sectors (financial services, logistics, procurement, customer service) are now agent- initiated rather than human-initiated, and is that share consistent with the acceleration argument in Part One, or slower? Over 2–5 years: does the automation-over-augmentation crossover that Anthropic observed on Claude.ai in 2025 replicate across other major model providers and across the wider economy, and if so, on what timetable does directive, minimal-oversight use become the default mode of AI interaction rather than the frontier case?

Signals and data to watch. Anthropic’s own methodology is the direct model: privacy-preserving classification of usage into automation versus augmentation modes, mapped onto a standard occupational taxonomy (O*NET, in Anthropic’s case) so results are comparable across time and geography, published on a fixed cadence (quarterly, in Anthropic’s V1–V3 releases) so trend, not just level, is visible. A state statistics office does not need proprietary conversation data to build an analogous instrument — it needs standing data-sharing arrangements with major model providers (on aggregated, privacy-preserving terms, following Anthropic’s own template), payment-rail data on machine-initiated transactions, and API-traffic proxies from cloud providers, none of which currently exist as a standing reporting requirement anywhere in this library’s source set.

Methods that fit. Time-series tracking against a fixed occupational or transaction taxonomy, published with methodology transparent enough that a rival institution could reproduce it — the opposite of a one-off survey. Cross-provider comparison matters more than single-provider depth, since a single lab’s usage index (however good) reflects that lab’s user base and cannot be assumed representative of the wider agentic economy.

Institutional wiring and first moves. A national statistics office or central bank research department should stand up an “agentic activity index” work stream within two quarters, explicitly modelled on the Anthropic Economic Index’s automation/augmentation taxonomy and O*NET mapping, and should approach at least two major model providers (not one, to avoid building an instrument that only ever reflects a single company’s user base) for aggregated, privacy-preserving data-sharing terms before the next AI Economic Index cycle makes doing so competitively awkward.

Priority 2 — Treat agent-platform concentration as a live competition-policy question now, not after the fact

In short. The regulators already investigating foundation-model concentration — the UK CMA, the FTC, the European Commission — are one layer too low. Agent-platform concentration is the next version of the same risk, and on the Cambridge paper’s evidence, it can move faster than the foundation-model layer did.

Why it ranks here. Report Two’s variable 5 already treats rent capture as a live risk grounded in real evidence: De Loecker-Eeckhout’s finding that average US markups rose from roughly 21% to 61% above cost since 1980 (NBER, Angle 06), and the superstar-firms literature tracing declining labour share to a small number of winner-take-most firms. The UK CMA’s technical update report on AI foundation models, the FTC’s 6(b) staff report on cloud-AI partnerships, and the European Commission’s competition-policy brief on industry concentration (all Angle 06 of this library) show regulators are already alert to concentration risk at the model layer. The Cambridge simulation’s finding — Gini coefficients moving from 0.24 to 0.70 depending on task-diversity design choices, winning bids compressing 14% under a single platform design decision — demonstrates that the agent layer sitting on top of that model layer carries the same concentration mechanism, mediated through platform rules (open versus sealed bidding, flat-fee versus performance-linked contracts) that almost no regulator currently has on its radar as a lever worth scrutinising.

The foresight questions and horizons. Near-term (this year): which firms are positioned to operate the dominant “agentic walled gardens” Microsoft Research’s paper describes — Apple, Google, Microsoft, Meta, OpenAI and Anthropic are all named as plausible operators by that paper’s own authors — and what interoperability commitments, if any, are being made or avoided as those platforms take shape? Medium-term (2–4 years): does the market converge on Microsoft Research’s “web of agents” (open, low switching costs, analogous to the early web) or “agentic walled garden” (closed, platform-controlled, analogous to today’s app stores) — and is that convergence happening through deliberate standard-setting or through unilateral platform lock-in that regulators notice only once switching costs have already hardened?

Signals and data to watch. Adoption and interoperability status of agent-to-agent protocols (Model Context Protocol, Agent2Agent) across major platforms; whether dominant consumer AI assistants restrict which service agents they can transact with (the “bowling-shoe” agent pattern the Coasean-singularity paper identifies, where a platform-provided agent enjoys privileged integration but limited portability); market-share concentration in agent-transaction volume by platform, tracked with the same rigour applied to foundation-model market share today.

Institutional wiring and first moves. Competition authorities already running foundation-model investigations (CMA, FTC, European Commission) should open a parallel, lighter-touch monitoring work stream on agent-platform interoperability now, rather than waiting for a market-power complaint to trigger a fresh full investigation — the lesson of the Cambridge paper being that by the time a complaint-driven investigation would normally open, a market of this speed could already have tipped. A standing information request to the major platforms on agent-to-agent interoperability commitments, modelled on the FTC’s 6(b) authority already used for the cloud-AI partnership report, is a low-cost first step available within a single regulatory cycle.

Priority 3 — Redesign labour-market safety nets for agent-speed displacement, not human-generation-speed displacement

In short. The reallocation-friction evidence Report Two draws on — Autor, Dorn and Hanson’s “China Shock” finding that trade-displaced workers took a decade-plus to reallocate (NBER, Angle 05) — describes a world where the competing force was other human workers and firms adjusting at human speed. The Cambridge paper shows agent-vs-agent competition can concentrate a market within the time horizon of a stylised simulation. Safety-net design built for the first kind of disruption will be too slow for occupations exposed to the second.

Why it ranks here. This is where Report Two’s variable 4 (labour reallocation friction) meets this report’s central finding most directly. The ILO’s refined global index of occupational exposure to generative AI, the OECD’s 2023 Employment Outlook chapter on AI and the labour market, and McKinsey Global Institute’s occupational-transition modelling (all Angle 05) were built to estimate how large a share of work is exposed and how long reallocation typically takes for workers displaced by automation of the traditional, tool-mediated kind. None of that literature was built with agent- vs-agent competitive dynamics in mind — dynamics the Cambridge paper shows can compress the effective disruption timeline for agent-exposed occupations from the years those models assume down to a period closer to a single retraining cohort’s enrolment window.

The foresight questions and horizons. Near-term: which occupations combine high exposure on the ILO’s index with high susceptibility to the kind of agent-vs-agent gig-platform competition the Cambridge paper models — freelance and platform-mediated work is the most obvious overlap, since it is structurally closest to the AI-Work simulation’s own setup. Medium-term: does reallocation time for workers displaced from agent-exposed occupations actually compress relative to the China Shock and Job Displacement and Job Mobility literature’s decade-plus benchmarks (both NBER, Angle 05), or does institutional friction on the human side of the labour market (licensing, geographic immobility, skills-matching delay) keep human reallocation slow even as the disruption that triggers it accelerates — producing a widening gap between how fast displacement happens and how fast re-employment can follow?

Signals and data to watch. Platform-level data on gig and freelance market concentration and pricing (the same variables the Cambridge simulation tracks — win rate, market share concentration, normalised bid price) in real gig-economy platforms, not just the simulated one; unemployment- duration and reallocation-speed statistics disaggregated by occupational exposure to agent-mediated competition, not just to automation exposure broadly, since the two are not the same variable and the existing ILO and OECD indices do not yet distinguish them.

Institutional wiring and first moves. Employment and welfare ministries should commission a follow-on to the existing ILO and OECD occupational-exposure work that specifically cross-references exposure to agent-mediated platform competition, not generic automation exposure — the distinction this report has drawn throughout. Wage-insurance and rapid-retraining voucher schemes, the standard policy response to the China Shock literature’s reallocation-friction finding, should be piloted on a compressed timeline (months, not the multi-year rollout typical of retraining programmes built against a decade-plus reallocation assumption) in the occupations where ILO exposure and platform- competition exposure overlap most, treating speed of activation as the design variable that matters most, ahead of programme scale.

Priority 4 — Use the Coasean-singularity logic to actively re-open institutional design

In short. If transaction costs are genuinely collapsing toward zero for agent-mediated activity, the state has an active choice in what forms in that space — not merely a defensive posture toward whatever the private sector builds first.

Why it ranks here. Part One’s second section argued that the Coasean logic determining the boundary of the firm is now, for the first time since Coase wrote in 1937, a live variable rather than a structural constant. The Coasean-singularity paper’s own examples of where agent adoption concentrates first — real estate, job search, freelance hiring, investment — are all markets where private-sector agent platforms are already forming. Public-service delivery and benefits administration are structurally identical markets by the same logic (high-stakes interactions, information asymmetry, repeated need to re-establish context with an unfamiliar counterparty) and are currently almost entirely unaddressed by any private or public agent infrastructure.

The foresight questions and horizons. Near-term: could agent-to-agent design reduce the transaction cost of benefits administration and eligibility determination — the paperwork and re-explanation burden the Microsoft Research paper identifies as the core friction agents dissolve — in the same way it is beginning to reduce it in real estate and freelance hiring? Medium-term: should public-service delivery build its own “service agents,” in the Microsoft Research paper’s vocabulary, that citizens’ own assistant agents can transact with directly, and if so, on what identity and accountability infrastructure, given Hadfield and Koh’s finding that no such infrastructure currently exists for AI agents anywhere?

Signals and data to watch. Pilot deployments of agent-mediated public-service interfaces in any jurisdiction (an early, concrete exemplar worth tracking directly is the UK’s Government Office for Science foresight function and its published work on AI in public administration, alongside Singapore’s Centre for Strategic Futures and Policy Horizons Canada — the three institutions this library treats as the closest working analogues for state-level agentic foresight capacity); uptake and complaint rates on any such pilots, since trust and inspectability — flagged as the critical constraint by the Coasean-singularity paper’s own authors — will determine whether citizens allow their assistant agents to transact with a state service agent at all.

Institutional wiring and first moves. A digital-government or public-service-delivery unit should commission a scoping pilot — modest, one or two benefit programmes, eighteen months — for an agent-to-agent interface that lets citizens’ own AI assistants query eligibility and submit applications programmatically, built explicitly on the identity-and-liability groundwork Hadfield and Koh flag as missing (a registered, accountable principal behind every transacting agent, at minimum), rather than waiting for a private vendor to define the standard the state then has to adopt.

Priority 5 — Build sovereign, public-interest agent capability

In short. Compute and chip sovereignty is already a recognised geopolitical fault line in this library’s Angle 13. The agentic layer is a new dimension of the same asymmetry — and one a state can still shape, because the agent layer is younger and less entrenched than the compute layer beneath it.

Why it ranks here. Angle 13 of this library documents the compute-access asymmetry in detail: the CSIS analysis of the 2024 US chip export-control tightening, RAND’s account of the resulting AI Diffusion Framework tiering countries by compute access, CSET Georgetown’s “Silicon Twist” tracking of chips still reaching restricted end-users despite controls, and the IMF’s “Mind the Gap” finding that AI’s growth effect could be more than double in advanced economies relative to low-income ones. Every one of those asymmetries recurs, in a faster and less mature form, at the agent layer: a state or firm without access to frontier-grade agents is not just slower at deploying AI tools — per DeepMind’s Virtual Agent Economies finding on unequal-capability negotiation, it is structurally disadvantaged in the agent-to-agent transactions its citizens and firms increasingly depend on, since “the more capable AI assistants tend to be more successful and negotiate better deals for their users.”

The foresight questions and horizons. Near-term: does frontier agent capability remain concentrated in the same handful of jurisdictions (principally the US, with China as the other compute-tiered bloc under RAND’s diffusion framework) that already dominate foundation-model training, or does the lower compute intensity of agent orchestration (relative to frontier model training itself) allow a wider set of states to build competitive sovereign agent capability even without frontier-model-scale compute? Medium-term: does agent-layer capability become a second, compounding axis of the World Bank’s “Beyond the AI Divide” concern — the same countries structurally disadvantaged on compute access falling further behind specifically because their citizens and firms transact through lower-capability agents in a market where capability asymmetry, per DeepMind’s finding, directly determines negotiated outcomes.

Signals and data to watch. National or regional sovereign-cloud and public-interest-AI initiatives extending their remit explicitly to agent orchestration and deployment, not just model hosting; procurement policy for any public-sector agent deployment (Priority 4’s pilots, for instance) specifying capability floors, to avoid the state itself becoming the disadvantaged party in agent-to-agent transactions with better-resourced private counterparties.

Institutional wiring and first moves. A digital-sovereignty or industrial-strategy ministry should extend any existing sovereign-compute or public-AI programme’s mandate explicitly to cover agent orchestration capability, not only model access, within the current planning cycle — treating “can our public and SME sector field agents competitive enough to negotiate on equal terms with counterparties running frontier-grade agents” as a distinct capability question from “do we have enough compute,” since DeepMind’s evidence suggests the two do not move in lockstep.

Priority 6 — The agentic engine ENSI itself would deploy

In short. Consistent with ENSI’s standard foresight-engine structure, the agentic-economy monitoring function should itself be built as a small set of named agent archetypes, each with a narrow mandate, feeding human judgement rather than replacing it.

• Scanning agents, tracking agent-transaction volume, task-delegation rates and the automation- versus-augmentation split across available public and licensed data sources — the standing instrument Priority 1 calls for, run continuously rather than as a periodic study, modelled on Anthropic’s own O*NET-mapped, privacy-preserving classification methodology.

• Scenario-simulation agents, stress-testing Report One’s five futures against each new quarter of incoming data from the scanning agents — not to pick a winner prematurely, but to flag when incoming evidence starts to favour one future over the others clearly enough to justify a policy response, and equally to flag when a scenario previously treated as low-probability starts moving.

• Market-structure early-warning agents, applying the Cambridge paper’s own diagnostic variables — market-share concentration (Gini-style), win-rate distribution, normalised bid-price trends — to real agent-platform and agent-labour-market data as it becomes available, precisely because those variables moved from 0.24 to 0.70 within a stylised simulation’s own short horizon, and a live early-warning system needs to be watching before, not after, a real market shows the same trajectory.

• Translation and briefing agents, converting the scanning and simulation layers’ output into the kind of short, decision-facing briefing a minister, a board or a regulator’s leadership can act on within a single reading — the standing failure mode this playbook is designed to avoid is not lack of data, it is data that never reaches a decision-maker in time to matter.

Humans retain judgement and accountability throughout: agents surface signal, simulate scenarios and draft briefings; the decision about what to do with any of it — where to intervene, what to regulate, what to fund — stays with the accountable human institution the agents serve, following the same accountability logic this report has argued the state needs to establish for every other agent deployed in the wider economy.

Priority 7 — First twelve months

A state or large institution acting on this report between now and mid-2027 should, concretely:

• By Q4 2026: commission the agentic-activity index work stream (Priority 1) and approach at least two major model providers for aggregated usage-data terms, modelled explicitly on Anthropic’s published Economic Index methodology.

• By Q4 2026: task the competition authority already running a foundation-model investigation (CMA, FTC or European Commission, depending on jurisdiction) with opening a lighter-touch parallel monitoring work stream on agent-platform interoperability, using 6(b)-style information requests as the low-cost first instrument.

• By Q1 2027: commission the follow-on occupational-exposure study (Priority 3) cross-referencing ILO and OECD exposure indices against platform-mediated competitive-displacement risk, with results due within two quarters given the compressed timeline the Cambridge paper’s evidence implies is appropriate.

• By Q2 2027: scope the public-service agent-to-agent pilot (Priority 4) — one or two benefit programmes, built on explicit identity-and-liability groundwork rather than deferred to a vendor’s default terms.

• By Q2 2027: extend the mandate of any existing sovereign-compute or public-AI programme to cover agent orchestration capability explicitly (Priority 5).

• By Q3 2027: stand up the first two agentic-engine archetypes — scanning and market-structure early-warning (Priority 6) — as a standing function reporting quarterly, timed to precede the next full-cycle refresh of this report series.

Closing: the option-value argument

Every element of this playbook has value even if Report One’s most likely outcome — modest, uneven growth, unevenly distributed across sectors and geographies — turns out to be exactly what happens, and even if the more dramatic agentic dynamics described in Part One never scale beyond the particular markets (freelance platforms, real estate, personal-finance search) where agent adoption is currently concentrated. A live measurement instrument is worth building whether or not the automation-over-augmentation crossover Anthropic observed on Claude.ai turns out to generalise, because a state without one is flying blind on the single fastest-moving input to its own growth forecast regardless of which way that input moves. A competition-policy monitoring function on agent-platform concentration is worth having whether or not the Cambridge paper’s winner-take-all dynamics materialise outside a stylised simulation, because the cost of building the capacity to look is low and the cost of not having it, if the dynamic does materialise, is a market that has already tipped before regulators start their first investigation. Faster-activating labour-market safety nets are worth having whether or not agent-vs-agent competition ever displaces workers faster than the China Shock’s decade-plus benchmark, because a safety net built for a faster shock still works perfectly well for a slower one, while the reverse is not true. Sovereign agent capability is worth building whether or not the compute-sovereignty asymmetries in Angle 13 turn out to be the dominant axis of AI-driven inequality, because the alternative — discovering after the fact that citizens and firms are structurally disadvantaged in agent-mediated negotiation and having no domestic capability to respond — is not a position any state should choose to be in by default.

This is, in the end, the same argument Hadfield and Koh make in their own closing lines, and it is the right note to end this series on: “where we end up within this vast space of possibility is a design choice.” Reports One and Two describe the range of futures and the variables that determine which one materialises. This report’s contribution is narrower and more operational: the fastest- moving, least-institutionally-prepared-for variable in that whole system is not how quickly a task gets automated. It is whether the counterparty on the other side of a growing share of the world’s economic transactions is still, in any meaningful sense, accountable to a human being at all. That is a question this generation of policymakers gets to answer directly, while the institutions that will determine the answer are still being built — which is exactly the moment foresight capability is worth the most, and exactly the moment it is cheapest to build.

There is a specific and uncomfortable fact about Czech Technical University that ought to organise everything it does next in this area. CTU’s two most-cited research topics of the last five years are “artificial intelligence” and “machine learning”. On five-year comparison it sits among the top five European research groups in computer vision and ninth in robotics. CIIRC coordinates ROBOPROX — 468 million CZK of Excellent Research funding — runs EDIH CTU and the AI-MATTERS Testing and Experimentation Facility, is a partner in two of the five European AI and robotics networks of excellence, and has an institute named after its founder built and opened in Jaipur. Since February 2026 the university has been led by a rector who is a professor of artificial intelligence and the founder of the AI Center at the Faculty of Electrical Engineering.

And in 2024, 31.8% of CTU’s first-year bachelor students failed — 51.1% at the Faculty of Mechanical Engineering, 48.1% at Transportation Sciences, 45.6% at Nuclear Sciences and Physical Engineering, 34.0% at Information Technology. The university’s institution-wide rulebook for artificial intelligence in teaching is a methodological instruction issued in September 2023, eight pages long, structured as a table of permitted, partly permitted and forbidden activities, written before the assessment-reform literature matured and before any of the randomised trials that now define the field had been published. Its named AI-literacy provision for students is a licensed online course produced by the University of Helsinki.

That gap — between what CTU knows about AI and what CTU does with AI in its own lecture theatres — is the whole strategic situation. It is not a criticism of the instruction, which was an early and sensible response and remains more concrete than most European universities managed. It is an observation that CTU is currently a world-class producer of artificial intelligence and an ordinary consumer of it, and that the second fact is now the more consequential one. Every technical university in Europe can license the same models. Almost none of them have a rector who built an AI research centre, a CIIRC-scale institute, a top-five European computer-vision group, and a national mandate under the Czech AI Strategy to 2030 sitting in the same building as a 51% first-year failure rate. The asymmetry is the opportunity, and it has a short half-life: the differentiation is available for about three years, after which everyone will have done the obvious things.

But attrition is only the first of two problems, and the second changes what the end of a degree is for.

Stanford’s Digital Economy Lab, tracking millions of payroll records, finds employment declining specifically among young workers in the occupations most exposed to AI — entry-level software and technical roles, which is to say the destination of a technical university’s bachelor and master graduates. Set that against the compression findings that make the tutoring case so strong: +34% for novices against near-zero for experts in the NBER field study of 5,179 workers; the largest gains for below-average performers in the BCG field experiment; +9 percentage points for students of the weakest tutors in Stanford’s Tutor CoPilot trial.

Put those together and the shape of the problem is unusual. The technology compresses the performance gap between novice and expert while eroding the jobs in which that gap was historically closed. A CTU graduate can now perform like a competent junior on day one, and may find no junior position in which to become a senior. The apprenticeship function — the first three years in industry where an engineer acquires judgement by being wrong under supervision — is migrating upstream, and the only institution positioned to absorb it is the university.

This is not a distant concern. It bears directly on what a capstone, a diploma thesis, an industrial placement and a laboratory course are for. CESAER’s Engineer of the Future white paper and the CDIO tradition already describe the direction — challenge-based, competence-based, real consequences — without naming this as the reason. Several of the highest-scoring ideas below are that argument made operational.

So the reframe this playbook argues for is not “CTU should adopt AI in teaching”. Everyone will. It is that CTU should treat teaching as an application domain of its own research, and its own students as the population on which the European evidence base gets built — against two problems at once: the third of the cohort lost in year one, and the professional formation that used to happen after graduation and no longer will. That is a claim about identity, not about procurement. A university that publishes on machine learning and runs its education on intuition and committee memory is holding two incompatible epistemologies.

Three conditions make the timing favourable, and all three are transient. The mandate already exists: Strategic Plan CTU 2021+ has four pillars, of which Pillar 1 is Study, and its goals include raising the quality and success rate of study and bringing practice into teaching; Pillar 3 commits CTU to “digitalisation of activities and operations, decision-making on the basis of data”. The evidence has arrived and points at CTU’s worst number, since every well-designed study says the benefit concentrates in the weakest performers. And the university has already run the experiment without noticing — the 2024 annual report records that the Faculty of Mechanical Engineering uses artificial intelligence to predict “at risk” status from weekly examination results and offers targeted help with exam scheduling, credited with a significant fall in failure. It is not in the strategic plan, has no institutional owner, and has never been evaluated to any standard CTU would accept from a doctoral student. The programme does not start from zero; it starts from an unowned success nobody has scaled.

How the ideas are scored

Forty interventions follow. Three dimensions, each out of 10, summed to a composite out of 30. A fourth consideration — feasibility — is deliberately kept out of the score and reported separately, because feasibility should determine sequence, not merit. Scoring hard things down because they are hard is how institutions end up with a portfolio of easy things that changed nothing.

Evidence quality (E) — how well the international library actually supports this. A 9 or 10 means multiple randomised or quasi-experimental studies converging, in comparable settings, on a learning or progression outcome. A 6 or 7 means consistent professional consensus, regulator guidance, or strong observational evidence from named deployments. A 3 or 4 means it is a reasonable inference from adjacent evidence but nobody has tested this thing. A 1 or 2 means it is a bet.

Depth (D) — how structural the change is. Does it alter the teaching production function, what a degree certifies, or who is accountable for what — or does it add a capability on top of an unchanged system? A 9 or 10 changes what CTU is for the student. A 5 or 6 changes how something is delivered. A 2 or 3 is an improvement that leaves every underlying structure intact. High depth is not automatically good; it measures leverage and risk simultaneously.

Expected outcome (O) — magnitude times probability, at CTU specifically. Not “is this a good idea in general” but “what is the realistic expected value of doing this at an institution with 18,168 students, 2,247 academic staff, 221 study programmes across eight faculties and six institutes, a 31.8% first-year failure rate, 19.7% international enrolment, and a rector who is an AI professor”. An idea with a large effect that will probably not survive institutional contact scores lower than a modest effect that certainly will.

Feasibility, reported separately as one of four labels: this year · this cycle (2–3 years) · accreditation-bound (lands at the next programme reaccreditation) · hard (requires external agreement, money that does not exist, or a cultural change with no current sponsor).

The composite is a ranking instrument, not an oracle. Two honest limitations. It rewards ideas whose effects are measurable, which biases against slow cultural moves that matter and cannot be counted. And the E scores are drawn from a library assembled for this question — a different library would move them. Both are reasons to read the reasoning rather than the number.

The scoreboard — forty ideas, ranked

Read as: E evidence quality · D structural depth · O expected outcome at CTU · = composite /30 · feasibility label.

Tier 1 — the seven that clear 24

1. Two-lane assessment at programme level — E 9 · D 9 · O 8 · = 26 · accreditation-bound. Replace MP 5/2023’s activity-by-activity permission schema with a small number of properly secured certification points per programme and everything else open and taught. Scores highest because the evidence is unusually complete on both halves — detection demonstrably fails, and TEQSA’s two-lane model has a five-year institutional track record — and because it changes what a CTU degree asserts, which nothing else on this list does as directly.

2. Gateway Tutor on the five worst first-year courses — E 9 · D 7 · O 9 · = 25 · this cycle. Constrained, course-grounded tutors on the gateway subjects at FS, FD and FJFI, run as randomised trials. Highest expected outcome on the list: four independent RCTs support the intervention, every compression study says the effect concentrates in exactly the population CTU is losing, and the target number is published annually.

3. Discipline-specific verification standards — E 8 · D 9 · O 8 · = 25 · this cycle. Define, per discipline, what counts as checking a machine-produced result — in circuits, in structures, in code, in control. Scores here because the jagged-frontier evidence shows people cannot locate the capability boundary untaught, and because this is the one graduate competence that is both load-bearing and currently taught nowhere.

4. The Junior Engineer Compact — E 7 · D 10 · O 8 · = 25 · hard. With Czech industry: a structured final-year track of supervised responsibility for real work, absorbing the apprenticeship function that entry-level roles no longer perform. The deepest idea on the list and the only one addressing the collapse of the junior pipeline. Marked hard because it requires industry agreement CTU does not yet have.

5. Four graduate AI outcomes in every accredited programme — E 8 · D 9 · O 8 · = 25 · accreditation-bound. Verification, frontier judgement, unassisted core reasoning, accountability for machine-produced results — written into programme documentation with named assessment points. Scores high on depth because accredited outcomes are the only teaching change that survives a change of dean.

6. Teaching redesign counts for promotion and habilitation — E 6 · D 10 · O 8 · = 24 · hard. The single highest-leverage move on the list and the one with the weakest direct evidence, which is why it sits sixth rather than first. Every study of stalled adoption identifies incentives; none of them tested changing incentives. Depth 10 because it is the only idea here that changes what the institution rewards.

7. Assessment twins — E 8 · D 8 · O 8 · = 24 · this cycle. Pair an open, AI-permitted component with a short secured oral or practical assessing the same outcome, scheduled close together for cross-verification. Scores just below two-lane assessment because it is the implementable pattern inside that model rather than the model itself.

Tier 2 — strong (20–23)

8. CS1/CS2 outcome rewrite at FIT and FEL — E 9 · D 7 · O 7 · = 23 · accreditation-bound. Rebuild introductory programming outcomes around decomposition, specification, verification, debugging and reading unfamiliar code. Best-evidenced curriculum change available; capped on outcome only because it touches two faculties.

9. A CTU-built, course-grounded tutoring layer — E 8 · D 8 · O 7 · = 23 · this cycle. Own the pedagogically-constrained layer rather than renting a general assistant. Georgia Tech’s 76.7%-versus-31.3% accuracy gap is the evidence that grounding, not model choice, is what works.

10. The Teaching Evidence Unit — E 9 · D 7 · O 7 · = 23 · this year. Three to four people who make every deployment a pre-registered trial. Highest evidence score on the list; outcome capped because its effect is entirely indirect — it changes the quality of every other decision rather than any student’s result.

11. ETH Zurich lecturer framework, workload-credited, by discipline — E 8 · D 7 · O 7 · = 22 · this year. Adopt rather than draft; deliver in faculty cohorts with hours credited, not added. Scores well and is capped by the honest observation that no study in the library evaluates a faculty AI-development programme against a teaching outcome.

12. The Confusion Map — E 6 · D 8 · O 8 · = 22 · this cycle. Publish, per course, where students actually get stuck, derived from the tutor interaction stream. A university has never before been able to see confusion at scale in the student’s own words; ten million CS50 queries is a map of what is hard about introductory computing that no pedagogical intuition could produce.

13. Retire detection from misconduct procedure — publicly, with the reason — E 9 · D 5 · O 7 · = 21 · this year. Fourteen detectors failed systematic testing; GPT detectors misflag roughly 61% of non-native English writers. With 3,577 international students and 78 English-taught programmes, this is a live equity exposure. Cheapest high-scoring move available.

14. Convert saved lecture hours into studio and seminar contact — E 6 · D 8 · O 7 · = 21 · this cycle. The productive use of every efficiency elsewhere on this list. Scores on depth because it is the difference between an agentic university and an automated one, and it will be decided in budget meetings rather than strategy documents.

15. Universal AI-delivered pre-matriculation bridge — E 7 · D 6 · O 8 · = 21 · this year. Scale what FIT, FEL and FJFI already run in fragments — FIKS, FEL Camp, the Preparatory Week, the Mathematical and Physics Minimum — into a universal, AI-delivered bridge for every admitted student. High outcome because it attacks the failure before enrolment, when it is cheapest.

16. University-wide Defence Week — E 6 · D 8 · O 7 · = 21 · this cycle. A scheduled institutional rhythm of oral authentication rather than per-course vivas invented by exhausted individual lecturers. Turns the staffing problem of secured assessment from a distributed impossibility into a timetabling exercise.

17. Redesign fellowships — E 7 · D 7 · O 7 · = 21 · this cycle. Pay academics released time to rebuild a specific course, with a deliverable. The Ithaka evidence is unambiguous that lack of time, not lack of willingness, is the binding constraint.

18. One governance register with a risk-class sequencing rule — E 8 · D 6 · O 7 · = 21 · this year. AI Act Annex III classification, ESG route, data flows and named owner for every system, plus a published rule that no high-risk deployment precedes an evaluated low-risk one. Protects the programme from its own enthusiasts.

19. Examiner calibration — E 6 · D 7 · O 7 · = 20 · this cycle. Use AI to measure and reduce inter-examiner variance. A larger and more measurable fairness problem than cheating, entirely unaddressed, and newly tractable.

20. Students build the agents — E 4 · D 9 · O 7 · = 20 · this cycle. FIT and FEL students build the tutoring agents for FS and FSv gateway courses as assessed coursework. Collapses the cost, produces an authentic capstone with a real user, and makes the university’s own teaching the object of student engineering. Evidence score is low because nobody has published this; depth is high because it changes who does the work.

21. Retain the interaction stream institutionally — E 6 · D 8 · O 6 · = 20 · this year. MP 5/2023 already states the fact correctly: no AI tool used at CTU is operated by CTU. Today the richest teaching-improvement dataset the university could own accrues to a vendor.

22. The AI-native capstone — E 5 · D 8 · O 7 · = 20 · accreditation-bound. Every capstone ships and publicly defends an artefact built with agentic tooling, assessed on design decisions and verification rather than authorship.

Tier 3 — worth doing (16–19)

23. Diagnostic competence map at entry — E 5 · D 7 · O 7 · = 19 · this cycle. Replace a single admission score with a per-topic gap map that routes the student to specific remediation.

24. “Machines and Judgement” spine course across all eight faculties — E 6 · D 7 · O 6 · = 19 · accreditation-bound. One shared, discipline-adapted course carrying the AI-literacy and verification content, replacing the licensed general online course currently doing that job.

25. The Course Concierge — E 8 · D 4 · O 7 · = 19 · this year. Syllabus and logistics agent. Low depth by design; it is the cheapest archetype, the fastest visible staff relief, and the safest place to learn to run any of this.

26. Teaching-AI as doctoral topics at CIIRC and the AI Center — E 5 · D 7 · O 7 · = 19 · this cycle. Solves staffing, cost and publication simultaneously by making the teaching engine a research programme rather than unpaid service.

27. The Week-Six Trigger — E 6 · D 5 · O 7 · = 18 · this year. Detect and remediate at the specific point in a cumulative course where recoverable falling-behind becomes unrecoverable.

28. Audit and scale the FS at-risk model — E 7 · D 4 · O 7 · = 18 · this year. It already runs, it is credited with a fall in failure, it has no institutional owner, and it has never been audited for cohort drift or subgroup fairness — which the dropout-prediction literature says is where these models fail.

29. Second-chance architecture — E 4 · D 7 · O 7 · = 18 · this cycle. A structured re-entry path with AI-supported catch-up for the near-miss share of the 25.71%, instead of treating failure as terminal.

30. The Removal Register — E 4 · D 8 · O 6 · = 18 · accreditation-bound. Require every programme to name what it retired this cycle. Curricula only ever accrete; adding AI content without a removal instrument produces an unteachable degree.

31. Embedded CIIRC engineer per faculty — E 4 · D 7 · O 7 · = 18 · this cycle. A semester-long residency that transfers capability rather than delivering a system and leaving.

32. Czech technical-language evaluation set — E 5 · D 7 · O 6 · = 18 · this cycle. Frontier models serve Czech technical instruction measurably worse than English. CTU has the NLP capability to build the benchmark, and no one else in the country will.

33. Write teaching-AI into the structural-fund proposals — E 4 · D 6 · O 8 · = 18 · this year. The AIML Research Centre and AI European Centre of Excellence are in preparation now. A workstream written in at drafting stage is funded; one added later is not. Pure timing value, and the window closes.

34. Process portfolio assessment — E 5 · D 7 · O 5 · = 17 · accreditation-bound. Assess the trajectory of work rather than the final artefact.

35. The cognitive gym — E 6 · D 6 · O 5 · = 17 · this cycle. Deliberately unassisted practice spaces, defended pedagogically rather than punitively — the capability on which verification skill is parasitic.

36. EuroTeQ workstream and EDIH industry microcredentials — E 5 · D 6 · O 6 · = 17 · this cycle. The export move. Worthless before delivery exists, valuable immediately after.

37. Validate the admission test against post-AI outcomes — E 5 · D 6 · O 5 · = 16 · this cycle. If AI compresses performance differences, the instrument that used to predict who succeeds may no longer predict it. Nobody has checked.

Tier 4 — scored, and not recommended now (≤15)

38. Assessment variant generation at scale — E 5 · D 4 · O 6 · = 15 · this year. Twenty equivalent exam versions, generated and human-checked. Genuinely useful — but it is a component of the secured lane, not an initiative, and promoting it to a programme invites building the tool before deciding the assessment model it serves.

39. Resit and exam-scheduling optimisation — E 5 · D 3 · O 5 · = 13 · this year. Real efficiency, no structural change, and it risks becoming the visible “AI project” precisely because it is easy and uncontroversial. Do it as operations, not as strategy.

40. Peer observation of AI-mediated teaching — E 5 · D 4 · O 4 · = 13 · this year. Sound practice, but with 2,247 academic staff it consumes exactly the senior attention the redesign fellowships need, and the evidence that observation changes teaching behaviour is weak.

Deliberately not on this list, and why. Automated summative grading, AI admissions triage and remote proctoring were considered and excluded rather than scored, because all three sit inside AI Act Annex III’s high-risk categories while CTU has no compliance track record, no evaluation function and no institutional trust built. They are not bad ideas permanently; they are bad ideas first, and the cost of that mistake is not recoverable.

Tier 1 in full

1. Two-lane assessment at programme level — 26/30

In short. Retire the activity-by-activity permission schema of Methodological Instruction 5/2023 and replace it with a programme-level architecture: a small, deliberately-chosen set of secured certification points where CTU asserts that a named human holds a capability, and everything else open, AI-permitted and taught.

The mechanism, and why it is not a rules change. The current instrument asks, of each activity, may a student use AI for this? That question has no enforceable answer, and the enforcement evidence is conclusive: Weber-Wulff and colleagues, working through the European Network for Academic Integrity, found fourteen detection tools neither accurate nor reliable and defeated by light paraphrase; Liang and colleagues at Stanford found GPT detectors misclassifying roughly 61% of essays by non-native English writers as AI-generated while performing near-perfectly on native speakers. With 3,577 international students — 19.7% of the body — from over 100 nationalities and 78 of 221 programmes taught in English, a detection-founded regime at CTU accuses its international cohort at several times the domestic rate on the basis of second-language fluency.

The two-lane model asks a different and answerable question: what is this assessment for? TEQSA’s 2023 discussion paper established the principles and its 2025 follow-up reports what institutions actually built from them. Assessment of learning certifies — and therefore requires secured conditions and identity assurance, at programme level rather than in every task. Assessment for learning develops — and there AI use is open, expected, and frequently the subject of the assessment. The unwinnable arms race came from demanding both jobs from every assignment.

Why it scores 26. Evidence 9: both halves are unusually well established — the failure of detection empirically, the two-lane model through five years of regulator-guided institutional practice. Depth 9: it changes what a CTU degree asserts to a stranger, which is the institution’s actual product. Outcome 8: high confidence of adoption because it reduces staff burden rather than adding to it, capped only by accreditation timing.

What CTU already has. More than most European universities. MP 5/2023 (ČVUT_MP_2023_05_V01, eight pages, effective 25 September 2023, issued by the Vice-Rector for Bachelor and Master Studies) already did the hard analytical work of thinking activity-by-activity about where AI use is pedagogically load-bearing — that analysis is reusable, it is the form that must change. Two of its provisions should be preserved verbatim: the warning that no AI tool used at CTU is operated by CTU and that all user–tool communication is visible to the operator, and the explicit treatment of deepfake identity modification in online examinations as a disciplinary offence. The Study and Examination Code, consolidated and effective from 1 December 2025, is the harder vehicle in which secured-assessment requirements must ultimately live.

The first move. Run the QAA four-step triage across every programme, through existing internal quality assurance rather than as an emergency parallel process — which is also what keeps it accreditable with NAÚ under the recommended procedures for preparing study programmes. For each programme the output is a single page: which outcomes require certification, at which points, under what identity assurance. Expect the honest answer to be three to five points across a bachelor’s degree, not thirty.

The failure mode. Two, both common. The first is that “secured” is implemented as surveillance — proctoring software, which is both an Annex III high-risk use and, on the evidence in this library on proctoring and disability, an accessibility liability. The second is that the open lane is declared and then quietly policed anyway, with informal suspicion migrating into marking. The countermeasure to both is measurement: publish the misconduct-rate disparity between international and domestic students annually.

Cost and owner. Vice-Rector for Studies, through faculty study committees and the Academic Senate. Drafting is cheap; the real cost is contact hours for secured oral components, which idea 16 (Defence Week) exists to make affordable. The Integrevise research report gives the staffing and cost model for oral assessment at cohort scale — compute it for CTU’s actual cohorts rather than dismissing the option on intuition.

The number that proves it wrong. If, two years in, no programme has reduced its number of assessed tasks and the misconduct disparity has not narrowed, this was a document change and not a reform.

2. Gateway Tutor on the five worst first-year courses — 25/30

In short. Constrained, course-grounded AI tutors on the gateway subjects that fail the most students — mathematical analysis, physics, mechanics, first programming — beginning at the Faculty of Mechanical Engineering, deployed as randomised trials.

The mechanism. First-year failure at a technical university has a stereotyped shape: the material is strictly cumulative, a student falls two weeks behind, week seven becomes unintelligible without week five, the only remediation is a consultation hour that clashes with another lecture, attendance stops, formal failure follows in February. Nothing in that sequence requires a human at the moment of intervention. It requires a patient, correct, course-specific explanation at eleven at night, which is the one thing this technology unambiguously supplies — provided it is built to withhold.

Why it scores 25. Evidence 9: four independent randomised trials converge — Kestin and colleagues’ Harvard physics crossover trial, where students learned more than twice as much in less time than in expert-led active learning; the World Bank’s six-week Nigerian RCT at 0.31 standard deviations; Stanford’s Tutor CoPilot at +4 points overall and +9 points for students of the weakest tutors; and a pooled meta-analytic effect of g = 0.670 across 35 studies and 4,193 participants. Outcome 9, the highest on the list, because the compression findings say the effect lands exactly where CTU’s losses are and the target metric is already published. Depth only 7 — it improves delivery of an unchanged curriculum, which is precisely why it is safe to do first.

What CTU already has. The Faculty of Mechanical Engineering already uses artificial intelligence to predict “at risk” status from weekly examination-period results and offers targeted help with exam scheduling — recorded in the 2024 annual report and credited with a significant fall in failure. CTU’s teaching-AI programme starts from an unowned success nobody has scaled. It also has a substantial existing scaffolding culture to attach to (FIKS, FEL Camp, the FJFI Preparatory Week and its free senior-student tutor system, FD’s mathematics and physics tutoring, the CIPS, ELSA and KC counselling centres), KOS and Moodle as substrate, and in CIIRC and the FEL AI Center the capability to build this properly.

The first move. Two courses, not twenty. Assemble each course corpus — lecture notes, problem sets, worked solutions, past examinations, rights-cleared textbook material — and build retrieval-grounded tutors over them with the pedagogical constraint specified as an engineering requirement: one step at a time, question before answer, never the final result to an assessed problem. Georgia Tech’s numbers are the argument for grounding over model choice: 76.7% answer accuracy against 31.3% for a generic assistant baseline on the same evaluation, and coverage climbing from 21% at 80% precision to over 96% at over 86% precision through iteration.

Then randomise. A waitlist crossover — half the cohort in semester one, half in semester two — is ethically clean, methodologically standard, and produces the single most valuable unpublished result in European educational AI: does a constrained tutor reduce first-year engineering attrition, and by how much. No study in this library answers that.

The failure mode. Instruction dilution, with a published number: CS50 reports 22% of ten million responses containing code blocks despite instructions not to give solutions, 48% at conversation level. Their fix was not a better prompt but teaching fellows reviewing and correcting behaviour. Budget the loop or buy the leak.

Cost and owner. Vice-Rector for Studies with build capability seconded from CIIRC or the FEL AI Center, course ownership retained by the department — never an IT project, which is the documented way this fails. Inference is trivial: CS50 ran at $1.50 per student per year. The real costs are roughly 25 hours of corpus preparation per course and a permanent fraction of a teaching-assistant post per course in supervision.

The number that proves it wrong. First-year bachelor failure at FS, currently 51.1%, against the randomised control. A programme that cannot move it has failed and should be said to have failed.

3. Discipline-specific verification standards — 25/30

In short. Define, write down and assess what it means to check a machine-produced result in each of CTU’s disciplines. Not “critical thinking about AI” as a generic disposition — the specific, technical, discipline-bound question of how an electrical engineer establishes that a circuit analysis is right, how a structural engineer establishes that a load path is right, how a programmer establishes that unfamiliar code does what it claims.

The mechanism, and why this is the load-bearing competence. Two findings define it. Dell’Acqua, Mollick and Lakhani’s field experiment with 758 BCG consultants found that inside the AI’s capability frontier participants produced work rated 40% higher in quality, while on a task just outside it they were 19 percentage points less likely to reach the correct answer than colleagues working with no AI at all — the tool did not merely fail to help, it degraded performance, because the boundary is invisible from inside and the failures are fluent. And Lee and colleagues at Microsoft Research and Carnegie Mellon, surveying 319 knowledge workers across 936 task examples, found that generative AI shifts effort toward verification and integration while higher confidence in the AI predicts less critical-thinking effort — the work moves to a task people are increasingly disinclined to perform.

For an engineer this is not productivity. A structural calculation, a control loop, a dosage algorithm, a safety interlock: being unable to tell that a plausible answer is wrong is how people are harmed. Verification is the competence on which professional liability rests, and it has never been taught explicitly because it used to be a by-product of doing the work by hand.

Why it scores 25. Evidence 8: the jagged-frontier and confidence findings are strong, replicated in shape across settings, and directly on point — though no study has yet taught verification and measured the result. Depth 9: it changes what the degree certifies. Outcome 8: high, because engineering already has the assessment forms to carry it and CTU’s disciplines are exactly the ones where “correct” is externally checkable.

What CTU already has. The single biggest structural advantage in this entire report and the one CTU under-uses: engineering assessment is already checkable against something that is not an examiner’s impression. A bridge calculation is checked by statics. A circuit oscillates or does not. Where humanities faculties must reconstruct authenticity from first principles, engineering mostly has to stop drifting away from it. CTU also has EUR-ACE/ENAEE and ABET-style outcome frameworks as the accreditation vocabulary, and the EuroTeQ Framework of Qualifications as the alliance-level architecture to write this into.

The first move. Commission each faculty to produce a two-page verification standard: the three to five checking procedures a graduate of that discipline must be able to perform, with worked examples of a fluent-but-wrong machine output in that domain and the procedure that catches it. Then assess it directly — give students tasks on both sides of the capability frontier and mark them on whether they knew which was which. That assessment is straightforward to build and almost nobody is building it.

The failure mode. Generic drift. The moment this becomes a university-wide module on “critical evaluation of AI outputs” it is worthless, because verification is not transferable across domains — checking a truss is nothing like checking a compiler optimisation. The countermeasure is that faculties write their own and are not permitted to adopt each other’s.

Cost and owner. Cheap in money, expensive in senior academic attention: this must be written by people who actually verify things professionally. Owner is each faculty’s study dean, coordinated by the Vice-Rector for Studies.

The number that proves it wrong. Every programme has a written verification standard with named assessment points within two accreditation cycles — or this was a memo.

4. The Junior Engineer Compact — 25/30

In short. With Czech industry: convert the final year of CTU’s engineering degrees into a structured track of supervised responsibility for real work with real consequences, formally absorbing the apprenticeship function that entry-level employment used to perform and increasingly does not.

The mechanism, and why this is the deepest idea here. The compression finding that makes idea 2 so strong contains a long-term problem that nobody has solved. AI raises the floor: +34% for novices against near-zero for experts in the NBER study of 5,179 workers; the largest gains for below-average performers at BCG; +9 points for students of the weakest tutors. Simultaneously, Stanford’s Digital Economy Lab finds employment declining specifically among young workers in AI-exposed occupations — precisely the entry-level technical roles where novices historically became experts by being wrong under supervision.

So the technology compresses the novice–expert performance gap while eroding the institution in which that gap was closed. A graduate performs like a competent junior on day one and may find no junior role in which to become a senior. Someone has to run the apprenticeship. The employer’s incentive to do it falls as the productivity gap between a graduate and an experienced engineer narrows. The university is the only remaining candidate.

Why it scores 25 with only E 7. Depth 10 — the maximum on this list — because it redefines what the last two years of an engineering degree are for: less coverage, more supervised responsibility, assessed on judgement under consequence rather than on completion. Outcome 8 on a long horizon. Evidence 7 because the diagnosis is well-evidenced (Stanford’s labour data, the compression studies, CESAER’s and CDIO’s challenge-based direction) while the intervention is not — nobody has run this and measured it.

What CTU already has. Deep industrial relationships and the institutional machinery to formalise them: EDIH CTU and the AI-MATTERS Testing and Experimentation Facility as industry channels; ROBOPROX; 380 partner universities; the existing diploma-thesis and industrial-project traditions, which are the seed of this and are currently assessed as documents rather than as professional performance. CESAER’s Engineer of the Future white paper — written by the association CTU belongs to — already argues for challenge-based learning without naming the collapsing junior pipeline as the reason. And strategic-plan goal 1.3, “bring practice into teaching”, is the existing mandate.

The first move. One faculty, one industrial partner, one cohort of twenty. Define what “supervised responsibility” means as an assessed outcome — the student owns a real deliverable with a real deadline and a real consequence, an industry engineer supervises, a CTU academic certifies the learning. The design question that must be answered first, and honestly, is what the firm gets: with AI compression, a final-year student supervised properly is genuinely productive, which is the argument to make rather than appealing to goodwill.

The failure mode. Two. It becomes an internship scheme — unstructured, unassessed, and indistinguishable from what already exists. Or it becomes free labour, which is an ethical failure and will be recognised as one. The guard against both is that the learning outcome is certified by CTU and the assessment is of judgement, not of output.

Cost and owner. Expensive in coordination, cheap in capital; the vice-rector responsible for cooperation with industry, jointly with a faculty willing to be first. Marked hard because it needs an industry agreement that does not currently exist, and because it will take an accreditation cycle to land properly.

The number that proves it wrong. Graduate employment and, more tellingly, time-to-first-independent-responsibility reported by employers. If graduates of the track are not measurably ahead within three years, the hypothesis was wrong.

5. Four graduate AI outcomes in every accredited programme — 25/30

In short. Compress the available competence frameworks into four assessable outcomes every CTU graduate must demonstrate, and require each of the 221 programmes to show where each is taught and where it is assessed — in the accreditation file, not in a strategy document.

The four, and why exactly these. Verification — can establish by independent means whether a machine-produced result is correct in their discipline (idea 3 supplies the discipline-specific content). Frontier judgement — can tell which side of the capability boundary a task sits on, which the BCG experiment shows people cannot do untaught. Unassisted core reasoning — can perform the discipline’s foundational reasoning without assistance, demonstrated at defined points, because verification is parasitic on it: you cannot check what you could never have derived. Accountability for results one did not personally generate — the professional stance of signing off on machine output, which every engineer already does with finite-element packages and library code, and which is now general.

Why it scores 25. Evidence 8: each outcome traces to specific findings rather than to aspiration. Depth 9: accredited learning outcomes are the only teaching change that survives a change of dean, a budget round, or the departure of the enthusiast who started it. Outcome 8: near-certain to persist once landed, discounted for the accreditation-cycle lag.

What CTU already has. The vocabulary exists and should be adopted rather than invented: UNESCO’s AI Competency Framework for Students (twelve competencies, four aspects, Understand / Apply / Create), the joint OECD–European Commission AI literacy framework, DigComp 2.2’s AI-specific knowledge and attitude examples, Digital Promise’s Understand / Evaluate / Use model, and the AI Literacy Heptagon’s translation into higher-education learning objectives. Structurally, CTU has the EuroTeQ Framework of Qualifications — the alliance deliverable defining what a European engineering graduate must be able to do — which is where these belong rather than in a parallel CTU-only scheme, and the NAÚ methodology for programme design as the accreditation route.

The first move. Draft the four outcomes at university level, then require every programme, at its next reaccreditation, to map them: where taught, where assessed, with what instrument. Do the mapping exercise on three pilot programmes first — one from FIT, one from FS, one from FA — because the honest result will be that most programmes cannot currently point to an assessment for any of the four, and it is better to discover that on three than on 221.

The failure mode. Documentation theatre — the outcomes appear in the file, nothing changes in the room. This is the standard fate of graduate-attribute schemes and it is worth naming in advance. The only reliable countermeasure is the assessment column: an outcome with no named assessment instrument is not an outcome.

Cost and owner. Almost nothing in money; considerable political effort across eight faculties; Vice-Rector for Studies with programme guarantors, using CTU’s EuroTeQ representation to push the same four upward so CTU is defining the alliance standard rather than adopting someone else’s.

The number that proves it wrong. Percentage of accredited programmes with all four outcomes mapped to a named assessment. If it plateaus below half, the outcomes were written at the wrong altitude.

6. Teaching redesign counts for promotion and habilitation — 24/30

In short. Change the criteria for promotion and habilitation so that substantial, evidenced teaching redesign counts as academic achievement — and make the evidence requirement real, so that it means an evaluated redesign rather than a claimed one.

Why this is the highest-leverage idea on the list. Every other proposal in this report is delivered by 2,247 academic staff who currently face an incentive structure that rewards publication and, at the margin, tolerates teaching. Ithaka S+R’s interview study across nineteen universities and its large-N national instructor survey describe the same picture: adoption in isolated pockets, driven by individual enthusiasm, blocked by frictions that have nothing to do with technology — no time, no recognition, no clarity, nobody to ask. The multi-institution barriers study shows the obstacles operating independently at individual, departmental and institutional level, which is the crucial finding: fixing time without fixing recognition changes little, because an academic who redesigns a course has spent a semester on something that will not appear in any file that decides their career.

Every one of the delivery ideas here — the tutors, the assessment redesign, the verification standards, the outcome mapping — is a large uncompensated ask of exactly the people whose promotion depends on something else. This idea is the one that changes the denominator.

Why it scores 24 with only E 6. Depth 10: it changes what the institution rewards, which is the deepest change available to any organisation. Outcome 8: if it lands, everything else gets easier; the discount is for the possibility that it is diluted into a box-ticking criterion. Evidence 6 is the honest constraint and the reason it ranks sixth rather than first — the literature identifies incentives as the binding constraint with great consistency, and not one study in this library tests changing them. That asymmetry should be stated rather than hidden: this is the best-diagnosed, least-tested intervention in the field.

What CTU already has. Strategic-plan Pillar 3 covers human resources, and goal 1.2 commits the university to raising the quality and success rate of study — the mandate exists. The habilitation and professorial-appointment framework is partly national and partly institutional, which bounds how far CTU can move alone; the internal promotion and evaluation criteria are entirely CTU’s.

The first move. Do not open the habilitation question first — it is the slowest and most contested. Start where CTU has unilateral control: internal performance evaluation, faculty-level promotion criteria, and the allocation of teaching-relief and institutional support. Define a recognised category of evidenced teaching redesign, with a specific evidentiary bar — a redesigned course, a pre-registered evaluation, a result, published. That bar is what stops the criterion degrading into “attended a workshop”, and it dovetails exactly with idea 10, the Teaching Evidence Unit, which supplies the evaluations.

The failure mode. Dilution into a checkbox, which is how most teaching-recognition schemes end. The countermeasure is the evidentiary bar and the fact that a published evaluation is externally legible in a way that a self-reported innovation is not.

Cost and owner. No direct cost, high political cost; the rector and the Academic Senate. This is the item that most requires the incoming rectorate’s authority, and it is worth noting that a rector who is an AI professor has unusual standing to argue that teaching with these systems is a serious intellectual activity rather than a service task.

The number that proves it wrong. The count of promotions in which evidenced teaching redesign was a material factor. If it is zero after two cycles, the criterion exists on paper only.

7. Assessment twins — 24/30

In short. For each outcome that must be certified, run two deliberately linked components: an open, AI-permitted piece of substantial work, paired with a short secured oral defence or practical demonstration of the same outcome, scheduled close together so each cross-verifies the other.

The mechanism. This is the implementable pattern inside idea 1’s architecture, and it resolves the practical objection that kills most two-lane implementations — that secured assessment at cohort scale is unaffordable. The insight is that the secured component does not have to carry the content; it only has to carry the authentication. A student who has genuinely done the open work can defend it in eight minutes. A student who has not, cannot, and no detector is required to establish that. The paper in this library builds the case through Messick’s validity framework: it shows precisely which assessment types generative AI undermines and why, then proposes the paired-component design as the response.

Why it scores 24. Evidence 8: the validity analysis is rigorous, the oral-assessment operating evidence is solid, and the design is a direct consequence of the two-lane principles that carry regulator backing. Depth 8: it changes the unit of assessment from the artefact to the artefact-plus-defence. Outcome 8: high confidence, because engineering already runs defences for theses and this generalises an existing form rather than importing a foreign one.

What CTU already has. The diploma-thesis defence is exactly this instrument, already operating at scale, already accepted culturally, already in the Study and Examination Code. The move is to generalise a form CTU already trusts down into coursework, not to invent one. The design studio’s crit is the same instrument in another register — and it is worth noticing that the Faculty of Architecture has by far the lowest first-year failure rate at 9.5%, in a faculty where the crit is relentless and work is defended continuously rather than submitted.

The first move. Pick the three or four courses per programme that carry the most certification weight and twin them. Then solve the timetabling, which is the actual constraint and is what idea 16 (a university-wide Defence Week) addresses: eight minutes per student per twinned assessment is impossible when every lecturer schedules it individually and entirely possible as an institutional rhythm. Compute CTU’s real numbers from the Integrevise oral-assessment cost model rather than dismissing it — for a 200-student cohort, one twinned assessment is roughly 27 examiner-hours, which is a scheduling problem, not an impossibility.

The failure mode. Inter-examiner variance. Oral assessment is only fair if examiners are calibrated, and they are typically not — which is why idea 19 sits adjacent to this one and should be done alongside it rather than after. The second failure mode is the defence degrading into a formality that everyone passes, at which point it certifies nothing and costs real hours.

Cost and owner. Contact hours, honestly stated in the workload model rather than absorbed silently by teaching staff — the single most common way assessment reform is quietly sabotaged. Faculty study deans, coordinated centrally for timetabling.

The number that proves it wrong. Pass-rate divergence between the open and secured components. If they agree perfectly, the secured component is not authenticating anything; if they diverge wildly, the open lane has a problem worth knowing about. Either result is informative, which is why this is worth instrumenting from the first cohort.

Tier 2 — the strong middle

These fifteen carry real weight and several are preconditions for Tier 1. Each gets the mechanism, the first move, and the catch.

8. CS1/CS2 outcome rewrite at FIT and FEL — 23. Mechanism: Becker and colleagues’ argument in Programming Is Hard — Or At Least It Used To Be is that many introductory objectives were proxies — the discipline never wanted students to write a loop from memory, it wanted decomposition, and the loop was how it checked. The proxy broke; the objective did not. Prather and colleagues’ observation of CS1 students using Copilot names the new pathologies precisely: drift, where the student’s mental model silently diverges from the accumulating code; over-trust of plausible output; collapse of the metacognitive loop. Ma, Chen and Konomi’s dialogue-log study adds that interaction pattern, not volume, predicts performance — and patterns are teachable. First move: rewrite outcomes around decomposition, specification, verification, debugging and reading unfamiliar code; adopt the ASEE pattern of requiring students to submit their own solution alongside the AI’s and account for the difference. The catch: FIT has 2,456 students and only six programmes — unusually redesignable — but assessment cost rises, because code critique cannot be autograded.

9. A CTU-built, course-grounded tutoring layer — 23. Mechanism: the pedagogy lives in the wrapper, not the model, and grounding is what produces accuracy — Georgia Tech’s 76.7% against 31.3% for a generic baseline. First move: define three explicit layers with owners — Microsoft Copilot under @cvut.cz accounts as the general floor (keep it; the Enterprise Data Protection posture is the hard part and is already solved), a CTU-built course-grounded tutor above it, and an institutional data layer above that. The catch: this only works if CIIRC and the FEL AI Center are commissioned as a research programme with publications and doctoral topics, not conscripted as an internal service desk.

10. The Teaching Evidence Unit — 23. Mechanism: the What Works Clearinghouse handbook and the EEF evaluator guide define what counts as evidence in education, and most published claims about AI in higher education would not qualify — they are satisfaction surveys without comparison conditions. First move: three to four people, reporting to academic governance rather than to the AI programme, with one non-negotiable rule — no deployment goes live without a comparison condition and a pre-registered outcome. The catch: it must not report to the people whose project it evaluates, or it produces evidence theatre. Its first task should be auditing the existing FS at-risk model, not blessing a new build.

11. ETH Zurich lecturer framework, workload-credited, by discipline — 22. Mechanism: the instruments already exist and are unused — UNESCO’s fifteen teacher competencies across five dimensions, DigCompEdu’s twenty-two competences on a six-level ladder, and ETH Zurich’s nine-page AI Competence Framework for Lecturers, built by a peer technical university for exactly this population. First move: adopt and translate ETH’s; map to DigCompEdu for European legibility; deliver one cohort per faculty, taught by academics of that faculty using that faculty’s own courses, hours credited against load. The catch: a mechanical engineer will not attend a generic session on prompt writing, and should not be asked to. Generic delivery is how this fails.

12. The Confusion Map — 22. Mechanism: the tutor interaction stream is a record of confusion at scale, in the student’s own words, at the moment it occurs — not inferred from exam performance months later. Ten million CS50 queries constitute a map of what is genuinely hard about introductory computing that no amount of pedagogical intuition could produce. First move: cluster and publish, per course, the top twenty points of difficulty each semester, to the course team first and the faculty second. The catch: it must be framed as course diagnostics, never as lecturer evaluation, or the data will be resisted and then gamed.

13. Retire detection from misconduct procedure, publicly — 21. Mechanism: fourteen detectors failed systematic testing; GPT detectors misflag roughly 61% of non-native English writers. First move: one paragraph in a rectoral instruction, withdrawing detector output as a basis for proceedings and stating the equity reason, so staff stop relying on it informally. The catch: announce it with the assessment triage from idea 1, or it reads as surrender rather than as reform.

14. Convert saved lecture hours into studio and seminar contact — 21. Mechanism: Georgia Tech’s 500-plus saved teacher hours are real, and they are overwhelmingly logistics and repeated questions. Whether that recovery becomes better teaching or becomes a staffing cut is the entire difference between an agentic university and an automated one. First move: commit in advance, in writing, that recovered hours are redeployed to contact and supervision — before the savings materialise and the budget round arrives. The catch: the supervision loop that makes the tutors work is exactly the line item that gets cut, and nothing visible breaks for several months.

15. Universal AI-delivered pre-matriculation bridge — 21. Mechanism: attack the failure before enrolment, when it is cheapest, and when the compression evidence says the marginal student gains most. First move: scale what already exists in fragments — FIKS at FIT, FEL Camp and the Embedded Technology Club, FJFI’s Preparatory Week plus its Mathematical and Physics Minimum, FSv’s mock entrance examinations, FD’s mathematics and physics tutoring — into one AI-delivered bridge offered to every admitted student, with a diagnostic front end. The catch: uptake is voluntary and self-selects for the students who need it least; the design problem is reach, not content.

16. University-wide Defence Week — 21. Mechanism: secured oral assessment is impossible when every lecturer schedules it alone and entirely possible as an institutional rhythm. First move: two fixed weeks in the academic calendar in which twinned assessments are examined, centrally timetabled, with examiner pools drawn across departments. The catch: it collides with everything else in the calendar and requires the Academic Senate to defend the slot against encroachment for at least three years before it becomes normal.

17. Redesign fellowships — 21. Mechanism: Ithaka’s evidence is unambiguous that time, not willingness, is the binding constraint. First move: twenty fellowships a year, one semester of teaching relief each, with a defined deliverable — a redesigned course, a documented assessment change, and an evaluation. The catch: without idea 6’s promotion criteria this rewards the already-committed and does not change the population; the two should be introduced together.

18. One governance register with a risk-class sequencing rule — 21. Mechanism: Annex III of the AI Act classifies as high-risk exactly the uses an efficiency-minded administration would automate first — admission, evaluation of learning outcomes, level assignment, examination monitoring — while tutoring, explanation and formative feedback are not on the list. First move: one register listing every AI system in teaching with its Annex III classification, provider-or-deployer status, data flows, human oversight, QA route and named owner; plus a published rule that no high-risk deployment precedes an evaluated low-risk one. The catch: modest work before the pilots, a project after the third.

19. Examiner calibration — 20. Mechanism: inter-examiner variance is a larger, more measurable and more consequential fairness problem than cheating, and it becomes critical the moment idea 7 puts more weight on oral assessment. First move: double-mark a sample, measure the variance, and use AI to generate calibration exemplars and to flag outlier marking patterns for human review — never to mark. The catch: it is politically delicate in a way detection never was, because the subject of measurement is academic staff rather than students.

20. Students build the agents — 20. Mechanism: FIT and FEL students build the tutoring agents for FS and FSv gateway courses as assessed coursework, supervised by CIIRC. Cost collapses, the capstone becomes authentic with a real user and a real evaluation, and the university’s own teaching becomes the object of student engineering. First move: one project cohort, one course, with the domain academic from the receiving faculty as the client. The catch: evidence 4 — nobody has published this, quality control is a genuine risk, and student-built systems must pass the same evaluation bar as any other before touching a real cohort.

21. Retain the interaction stream institutionally — 20. Mechanism: MP 5/2023 states the fact correctly — no AI tool used at CTU is operated by CTU, and all user–tool communication is visible to the operator. Today the richest teaching-improvement dataset the university could own accrues to a vendor. First move: make institutional retention of interaction logs a requirement of the layer-two build, under the Jisc Code of Practice for Learning Analytics. The catch: it only becomes an asset if someone is funded to read it, which is idea 12.

22. The AI-native capstone — 20. Mechanism: every capstone ships and publicly defends an artefact built with agentic tooling, assessed on design decisions, verification and defence rather than authorship — the CDIO project-based rubric pattern with explicit AI-use criteria. First move: rewrite the capstone rubric in one programme and run it for a year before generalising. The catch: team projects can be substantially machine-generated in ways a busy assessor will not detect, so the defence carries the assessment weight, not the artefact.

Tier 3 — worth doing, in the slipstream of something bigger

These fifteen are real but should not carry a programme. Most become cheap once a Tier 1 or Tier 2 item has built the capability they depend on.

23. Diagnostic competence map at entry — 19. Replace a single admission score with a per-topic gap map generated from a diagnostic, routing each admitted student to specific remediation. It is the front end that makes the pre-matriculation bridge (15) land on the students who need it rather than on the ones who volunteer. Cheap once the bridge exists; pointless before it.

24. “Machines and Judgement” spine course — 19. One shared, discipline-adapted course carrying AI literacy, verification and professional accountability across all eight faculties — replacing the licensed general online course (CTUPRGEAI, “Elements of AI”) currently standing in for this. Scores moderately because a course is a weaker instrument than an outcome: idea 5 does the durable work, and this is where it gets taught.

25. The Course Concierge — 19. A syllabus-and-logistics agent, built on the SyllabusQA pattern with an explicit factuality metric. Depth 4 by design and that is the point: it is the cheapest archetype, gives the fastest visible relief to teaching staff, and lets the institution learn to run one of these where the cost of an error is a corrected deadline rather than a corrupted understanding of thermodynamics. Do it first in time, not first in importance.

26. Teaching-AI as doctoral topics at CIIRC and the AI Center — 19. Makes the engine a research programme rather than unpaid service: doctoral students build and evaluate the tutors, publish the results, and the cost, the staffing and the evidence problem are solved by the same move. Depends entirely on the receiving academics treating it as real work, which is idea 6 again.

27. The Week-Six Trigger — 18. Cumulative courses have a specific point at which recoverable falling-behind becomes unrecoverable, and it is identifiable per course from historical data. Detect it and intervene there rather than at the examination. A refinement of idea 2 rather than a separate programme, and it needs the interaction stream to be worth much.

28. Audit and scale the FS at-risk model — 18. It already runs, the annual report credits it with a significant fall in failure, it has no institutional owner, and nobody has checked it for cohort drift or subgroup fairness — which is precisely where the dropout-prediction literature says these models fail, and fail worst for the groups an institution most wants to help. Low depth, high urgency: this is an unaudited model running on real students today.

29. Second-chance architecture — 18. Total study failure was 25.71% in 2024, and a large share of that is near-miss rather than incapable. A structured re-entry path with AI-supported catch-up converts some of it back. Scores modestly on evidence because nobody has evaluated such a path with AI support, and high on depth because it changes whether failure is terminal.

30. The Removal Register — 18. Require every programme, at each revision, to name what it retired. Curricula only accrete; adding AI competence to 221 programmes without a removal instrument produces a degree nobody can complete. Costs nothing, is universally resisted, and is the quiet precondition for ideas 5, 17 and 24 not making things worse.

31. Embedded CIIRC engineer per faculty — 18. A semester-long residency inside a faculty rather than a system delivered and abandoned — capability transfer, not procurement. Cheap, and it addresses the departmental level of the three-level adoption barrier that neither training (individual) nor policy (institutional) reaches.

32. Czech technical-language evaluation set — 18. Frontier models serve Czech technical instruction measurably worse than English, and no one has quantified how much worse in mechanics, structures or circuits. CTU has the NLP capability, the sovereign interest is real, and 143 of its 221 programmes are not in English. A small, publishable, genuinely national asset.

33. Write teaching-AI into the structural-fund proposals — 18. The AIML Research Centre and the AI European Centre of Excellence are in preparation. A workstream written in at drafting stage is funded from structural funds; the same workstream added afterwards comes out of the teaching budget. Pure timing value, zero intellectual content, and the window is closing now — which is why it appears in the first ninety days of the sequence despite scoring 18.

34. Process portfolio assessment — 17. Assess the trajectory of work rather than the final artefact. Pedagogically strong and administratively heavy; it is the right instrument for design and thesis work and the wrong one for a 400-student service course.

35. The cognitive gym — 17. Deliberately unassisted practice spaces, defended pedagogically rather than punitively — the capability on which verification skill depends, since you cannot check what you could never have derived. Scores modestly because it is a framing of the secured lane rather than an independent intervention, and because it is easily caricatured as nostalgia.

36. EuroTeQ workstream and EDIH industry microcredentials — 17. The export move: propose the teaching-AI programme as a EuroTeQ deliverable across an alliance of 115,000 students, and package the faculty-development programme as a paid microcredential for Czech industry through EDIH CTU. Worthless before delivery exists — announcing leadership before building capability is the most reliable way to discredit the whole programme — and valuable immediately afterwards.

37. Validate the admission test against post-AI outcomes — 16. If AI compresses performance differences, the instrument that historically predicted who succeeds at CTU may have quietly stopped predicting it. Nobody has checked, the data exists, and the answer matters for every recruitment decision. It ranks low only because it is a study rather than an intervention — but it is a cheap study that the Evidence Unit could run in its first year.

The portfolio view — what depends on what

A ranked list is not a plan, because several high scorers are useless without a low scorer underneath them. Four dependency chains govern the sequence.

The assessment chain. Retiring detection (13) is the precondition for the two-lane triage (1), which defines the certification points that assessment twins (7) implement, which are only affordable if Defence Week (16) exists and only fair if examiner calibration (19) runs alongside. Four of those five are Tier 2 or lower, and the Tier 1 item at the top of the chain cannot land without them. This is the clearest case in the report of ranking and sequencing pulling apart.

The delivery chain. The Course Concierge (25) builds the corpus pipeline, the logging architecture and the institutional confidence that the Gateway Tutor (2) then needs. The tutor produces the interaction stream (21), which becomes the Confusion Map (12), which is what makes the whole thing improve rather than merely run. Skipping the Concierge to get to the tutor faster is the commonest and most expensive shortcut available.

The capability chain. Nothing in the delivery or assessment chains is executed by anyone other than the 2,247 academic staff. The lecturer framework (11) supplies the competence, the redesign fellowships (17) supply the time, and the promotion criteria (6) supply the reason. All three are required; any two of them produce a programme that runs on volunteers and stops when they tire.

The evidence chain. The Evidence Unit (10) is upstream of everything, because it is what converts activity into knowledge — and it is also what makes the promotion criteria (6) meaningful, since “evidenced teaching redesign” requires someone to produce the evidence. Standing it up late means the first year’s deployments are unevaluable forever; once every student has the tutor, the counterfactual is gone.

The practical consequence: the first ninety days should be dominated by Tier 2 and Tier 3 items, and that is not a compromise. It is what the dependency structure says.

Thirty-six months

Days 1–90 — the free moves and the closing windows. Retire detection publicly with the equity reason (13). Name the accountable vice-rector. Stand up the Evidence Unit (10) and build the AI Act register (18). Write the teaching-AI workstream into the AIML Centre and AI European Centre of Excellence proposals (33) — this is the only item on the list with a hard external deadline. Commission the Course Concierge (25) and begin corpus preparation for two gateway courses. Audit the FS at-risk model (28), because it is running unaudited today. Adopt and translate the ETH Zurich lecturer framework (11).

Months 4–12 — first delivery and the incentive question. Gateway Tutor version one live in one course as a randomised waitlist trial (2). First discipline cohort of the lecturer programme at FS — the faculty with both the worst failure rate and the strongest incentive to move. Two-lane triage completed for FIT and FEL (1). First twenty redesign fellowships awarded (17). Open the promotion-criteria question (6) — it takes the longest and should start now, at internal evaluation and faculty promotion level where CTU has unilateral control, not at habilitation. Faculties commissioned to draft verification standards (3).

Months 12–24 — depth. Second and third gateway courses; first trial published. Assessment twins piloted in three or four high-certification courses (7), with the first Defence Week timetabled (16) and examiner calibration running alongside (19). Four graduate outcomes drafted and mapped on three pilot programmes (5). CS1/CS2 rewrite at FIT enters accreditation (8). Confusion Map published to course teams (12). Junior Engineer Compact scoped with one faculty and one industrial partner (4).

Months 24–36 — structure. Graduate outcomes into the accreditation pipeline across faculties. Verification standards written and assessed. Junior Engineer Compact running with its first cohort of twenty. Promotion criteria in force. EuroTeQ workstream proposed and industry microcredentials launched (36) — last, because the export is only credible once the delivery exists.

The seven numbers

Everything above reduces to seven indicators, five of which CTU already publishes or could publish tomorrow in “ČVUT v číslech”.

• First-year bachelor study failure, by faculty — 31.8% university-wide, 51.1% at FS, 48.1% at FD, 9.5% at FA. The primary outcome. Every other number here is instrumental to this one.

• Misconduct-rate disparity, international versus domestic students — measurable today, almost certainly uncomfortable, and the honest test of whether the assessment reform is real or cosmetic.

• Programmes with documented secured certification points — currently zero of 221.

• Academic staff at a defined DigCompEdu-mapped level, by faculty — currently unmeasured, which is itself the finding.

• Trials completed with a control condition and a pre-registered outcome — currently zero. This is the number that decides whether CTU cites others’ evidence for the next decade or generates it.

• Promotions in which evidenced teaching redesign was a material factor — currently zero, and the single best indicator of whether anything structural changed.

• Time to first independent professional responsibility, reported by employers — the long-horizon test of the Junior Engineer Compact, and the only one that measures whether the degree still does what it claims.

Close — what the scoring actually showed

Three things fell out of scoring forty ideas that did not fall out of choosing ten.

The highest-leverage idea is the worst-evidenced one. Changing what promotion rewards (6) scored a 10 on depth and a 6 on evidence, because the literature identifies incentives as the binding constraint with remarkable consistency and not one study has tested changing them. That asymmetry is worth sitting with. It means the field’s most confident diagnosis is also its least tested prescription, and an institution that acted on it would be doing something genuinely novel rather than following practice. For a university that already publishes on machine learning, running that experiment properly — and measuring it — is a more interesting contribution than another tutoring pilot.

Ranking and sequencing came apart, and the sequence should win. The dependency chains put Tier 2 and Tier 3 work in the first ninety days and leave the highest-scoring item, two-lane assessment, waiting on an accreditation cycle. An institution that executes strictly in score order will stall, because it will attempt the deep structural moves before the cheap enabling ones exist. The score says what matters; the chain says what is possible next.

The second problem is as large as the first, and far less worked on. Attrition is the largest measurable loss and idea 2 is the most reliable way to attack it — that much is settled by the evidence. But the collapse of entry-level technical employment puts an equally large question over the end of the degree, and unlike attrition it has no established answer anywhere. Only one idea in forty attempts one, and it scored a 10 on depth for exactly that reason. A list of forty options that contains a single response to a problem this size is itself a finding about the state of the field.

Which leaves the reframe the whole exercise keeps returning to: CTU does not need to become a different kind of institution to do any of this. It needs to apply to its own teaching the standard of evidence it already demands of its doctoral students — and to reward the people who do it. The first is a method the university already owns. The second is a decision only the rectorate can make, and on the evidence assembled here it is the one that decides whether the other thirty-nine ideas are a programme or a document.

Almost every university discussion of artificial intelligence is a discussion about a chatbot. Should students use one, may staff use one, which one should we license. This is a category error of a familiar kind — the same error as asking, in 1995, whether the university should have a website. The website was not the change. The change was that every administrative and scholarly process the institution ran would eventually be rebuilt around networked access, and the organisations that treated the web as a publication channel spent a decade being slowly reorganised by the ones that treated it as a substrate.

Teaching is not a single activity that a tutor-bot either helps with or does not. It is a production system with a dozen distinguishable functions, each with its own inputs, its own quality criterion and its own cost structure. Someone designs the course. Someone explains the material. Someone answers the question that gets asked forty times a semester. Someone marks the formative work. Someone notices that a student has stopped attending. Someone certifies, under their professional signature, that a named human can do a thing. Someone decides whether any of it worked. Those are different jobs, and generative AI has a radically different relationship to each of them. Some it can absorb almost entirely; some it can amplify by an order of magnitude; some it must be kept out of, for reasons that are pedagogical, legal, and — in the case of certification — constitutive of what a university is for.

The unit of design, therefore, is not a chatbot. It is an operating model: which functions are run by agents, which by humans, which by humans supervising agents, what data flows between them, and who is accountable when the whole thing produces a wrong answer to a student at midnight. This report specifies that model.

The evidence that it is buildable is no longer speculative, and it is unusually honest. Georgia Tech has run Jill Watson across its online master’s programme for years and published the operational numbers: more than 4,000 students across more than a dozen classes, more than 500 teacher hours saved, question coverage rising from about 21% at 80% precision in 2017 to over 96% coverage at over 86% precision by autumn 2019, and a companion system that generates a new agent for a fresh syllabus in about 25 hours. The LLM-era rebuild reports 76.7% answer accuracy against 31.3% for a generic assistant baseline on the same evaluation — which is the single most important architectural finding in this library, because it says the value is in the grounding, not the model. Harvard’s CS50 has served approximately 211,000 students and 10 million queries at $1.50 per student per year, with 94% of students finding the tools helpful. These are not pilots. They are production systems with published failure rates.

And the failure rates are the reason this report is structured around an operating model rather than a shopping list. CS50’s own evaluation reports that roughly 2.1 million of those 10 million responses — 22% — contained code blocks despite the system being instructed not to give solutions, rising to 48% at conversation level. The team names the mechanism: instruction dilution, in which a long conversation progressively erodes the authority of the system prompt until the pedagogical constraint simply stops binding. Their remedy was not a cleverer prompt. It was a human-feedback loop with teaching fellows reviewing and correcting the agent’s behaviour. Which means the working architecture was never “an agent”. It was an agent plus a staffed quality process, and every institution that budgets for the first and not the second is buying the leak without the benefit.

There is a second argument for the agentic frame, and for a technical university it is the more important one. The profession these students are entering is being rebuilt on exactly this substrate. The GitHub–Microsoft–MIT randomised trial found developers completing tasks roughly 56% faster with an AI pair programmer. The NBER field study of 5,179 support agents found +34% for novices and near-zero for experts, with the mechanism being diffusion of expert tacit knowledge. The BCG field experiment found consultants inside the AI frontier producing work rated 40% higher and, on a task just outside it, 19 percentage points less likely to be correct than colleagues with no AI at all. Whatever an engineering degree certifies in 2030, it will be exercised by a person who works by specifying, directing, verifying and integrating the output of machine systems. A university that teaches about that while teaching by nineteenth-century means is teaching the content and withholding the method.

So the reframe: an agentic university is not one that has bought AI. It is one whose teaching runs on the same substrate as the practice it certifies — deliberately, visibly, and with the human accountabilities specified rather than assumed. The students learn the discipline and, at the same time, learn what it is to work in a system where a competent machine does a large share of the first draft and a human is answerable for the result. That second lesson is the graduate attribute the labour market will actually pay for, and it cannot be delivered by a module.

One caution before the architecture, because the enthusiasm in this field is dangerous. The complementarity literature — Hemmer and colleagues’ review formalising when a human–AI team beats either alone — reports an empirical record that is frequently disappointing: many studies find the team underperforming the better of its two members. Complementarity is not automatic; it is engineered, and it is engineered mostly by deciding correctly where the boundary sits. Most of the design effort in what follows is boundary work.

The engine in brief

• Eight agent archetypes do the work, and they are not interchangeable: the Gateway Tutor (constrained, course-grounded, the highest-value one); the Course Concierge (logistics, cheapest to build, immediate staff relief); the Feedback Agent (formative only, never summative); the Lab and Simulation Agent (digital twins where physical capacity is the constraint); the Studio Critic (critique, never generation); the Instructional Design Agent (drafts syllabi, slides and assessment for human authorship); the Early-Warning Agent (triggers offers of help, never decisions); and the Evidence Agent (runs the trials and audits the others).

• Four data strata carry the system: the course corpus, the interaction stream, the progression record and the outcome ledger. The interaction stream is the asset nobody is capturing and the one that compounds.

• Four human roles must exist by name, or the architecture degrades: the course owner, the agent steward, the evidence lead and the certifying academic. The last one is not delegable — not culturally, but legally.

• The graduate this produces is an engineer defined by four capabilities the machine does not have: verification, frontier judgement, unassisted core reasoning, and accountability for a result they did not personally generate.

• Six failure modes are documented rather than hypothetical: instruction dilution, model drift across cohorts, vendor dependency, evidence theatre, fairness drift, and the hollowing of the apprenticeship — the last being the one nobody has solved.

How this report is organised

The architecture is presented in four layers, from the visible to the structural. First the agent archetypes, each with what it does, the published evidence that it works, its hard boundary — the thing it must never be allowed to do — and its human counterpart. Then the data layer, because agents without grounding are the failure case the Georgia Tech numbers quantify most starkly. Then the operating model: the named human roles and the accountability structure, which is where every deployment in the literature actually succeeds or fails. Then the graduate profile and the failure modes, which are the same subject viewed from opposite ends.

Two rules govern the whole design and are worth stating before the detail, because every archetype below is an application of them. Rule one: agents may generate, explain and flag; humans certify. The EU AI Act’s Annex III makes this a legal boundary as well as a professional one, classifying as high-risk any system used to evaluate learning outcomes, determine admission, assign educational level, or monitor examination behaviour. Rule two: every agent is grounded in institutional material and instrumented from the first day. An agent that answers from the model’s general knowledge is a liability; an agent whose behaviour is not logged and evaluated cannot be improved, defended, or shut down on evidence.

The eight archetypes

1. The Gateway Tutor — the constrained explainer

What it does. Answers a student’s question about the material of a specific course, at any hour, grounded in that course’s own lecture notes, problem sets, worked solutions and past examinations — and does so by withholding. One step at a time; a question back before an answer forward; never the final result to an assessed problem.

The evidence. This is the archetype with the strongest support in the entire library, and the support is unusually clean. Kestin and colleagues’ randomised crossover trial in Harvard’s PS2 physics course found students learning more than twice as much in less time with a pedagogically-constrained AI tutor than in an expert-led active-learning class — the comparison being against the best-evidenced form of human undergraduate physics teaching, not against a bad lecture. The World Bank’s six-week Nigerian RCT returned 0.31 standard deviations. The pooled meta-analytic effect across 35 experimental studies and 4,193 participants is g = 0.670. And the pre-LLM baseline — Ma, Adesope, Nesbit and Liu’s meta-analysis across 107 effect sizes and 14,321 learners — establishes what intelligent tutoring systems achieved before: real gains over large-group instruction, but not over individual human tutoring. The current generation is producing larger effects in weeks, in domains nobody hand-authored.

The hard boundary. It must not solve assessed work, and the constraint must be treated as an engineering requirement with a measured failure rate rather than a line in a prompt. CS50’s published 22% code-leakage rate is the number to design against, and instruction dilution is the named mechanism: in long conversations the system prompt loses authority. Practical countermeasures visible in the deployments — conversation-length limits and re-anchoring, retrieval that returns the scaffolding material rather than the solution, refusal classifiers running as a separate check rather than as prompt text, and sampled human review.

Its human counterpart. A named teaching assistant per course reviewing sampled conversations weekly. This is the CS50 lesson stated as an organisational requirement: the human-feedback loop was the fix, not the prompt.

Where it goes first. The course with the worst failure rate, because every compression finding in this library — NBER’s +34% for novices, Tutor CoPilot’s +9 percentage points for students of the weakest tutors — says the effect concentrates at the bottom of the distribution.

2. The Course Concierge — logistics and syllabus

What it does. Answers the question that is not about the material: when is the deadline, what is the resit policy, which room, what counts toward the grade, is the lab report due before or after the exam. This is a startlingly large share of what teaching staff actually spend time on and it carries no pedagogical value whatsoever.

The evidence. The SyllabusQA work in this library is the reference — 63 syllabi, 5,078 question–answer pairs, retrieval-augmented baselines and, importantly, a factuality metric, because a confidently wrong answer about an examination date is worse than no answer. Georgia Tech’s operational figure belongs here too: Jill Watson’s 500-plus teacher hours saved across a dozen classes is overwhelmingly this category of question, not deep conceptual tutoring.

The hard boundary. It must never invent a policy. The correct failure behaviour is escalation with an explicit “I do not have this in the course documents”, which requires the retrieval corpus to be authoritative and current — meaning someone owns keeping it current.

Its human counterpart. The study office, which stops answering the same forty questions and starts answering the hard ones.

Why it goes first in practice. It is the cheapest archetype to build, it has the fastest visible payoff for staff, and — the political point — it builds institutional trust in the technology on a use case where the downside of an error is a corrected deadline rather than a corrupted understanding of thermodynamics. Every deployment sequence in the literature that succeeded started somewhere low-stakes.

3. The Feedback Agent — formative only

What it does. Gives a student a substantive response to a draft, a solution attempt, a design or a piece of code — against the actual rubric — before it is submitted for a grade. Not a mark. A critique.

The evidence. The computing-education literature carries the strongest signal: the ITiCSE working group’s survey covers automated feedback extensively, and the pattern documented in the ASEE study — students submitting their own solution alongside the AI’s and accounting for the difference — converts feedback into the object of study rather than a substitute for effort. The EducationQ benchmark, a multi-agent teacher/learner/evaluator evaluation over 1,498 questions across 13 disciplines, supplies the most useful and least comfortable finding for procurement: teaching ability does not scale linearly with model size. The biggest model is not automatically the best teacher, which means model selection has to be evaluated on teaching behaviour rather than assumed from benchmark scores.

The hard boundary. It must not grade. This is the sharpest line in the whole architecture and it is drawn in law, not only in pedagogy: Annex III of the AI Act classifies systems used to evaluate learning outcomes as high-risk, with the full apparatus of risk management, documentation, logging and human oversight attaching. A formative agent that quietly becomes the basis of a mark has silently reclassified itself and the institution’s obligations along with it.

Its human counterpart. The academic who sets the rubric and owns the summative judgement. In practice, the useful arrangement is the Tutor CoPilot pattern — the agent coaches the human marker rather than the student, which is the configuration that produced measured learning gains at roughly $20 per tutor per year.

4. The Lab and Simulation Agent — where physical capacity is the constraint

What it does. Runs guided experimental work in simulation: a digital twin of the plant, the circuit, the structure or the process, with an agent that sets the task, watches the student’s actions, asks why, and injects the fault that makes the lesson land.

The evidence. The gAI-PT4I4 framework in this library is the strongest published template — generative AI combined with low-fidelity digital twins, VR and retrieval-augmented generation for personalised experiential learning in an Industry 4.0 training context. The AITEE agentic tutor for electrical engineering is the discipline-specific pattern: graph-based retrieval over course content plus tool use, applied to circuit analysis, which is exactly the shape of problem a technical university has a thousand instances of.

The hard boundary. Simulation must not silently replace the physical laboratory. An engineer who has only ever debugged a simulation has not encountered the thing that makes engineering difficult — that the model is not the world, that the connector is loose, that the tolerance stack-up ate the margin. The correct framing is capacity extension: unlimited rehearsal before scarce bench time, and unlimited variation afterwards, so the physical session is spent on what only physical presence teaches.

Its human counterpart. The lab supervisor, whose time is redirected from setup and standard procedure toward the moments where physical judgement is formed.

5. The Studio Critic — critique, never generation

What it does. Sits with a design student — architecture, mechanical design, systems design — and interrogates the design: what is this load path, why this material, what happens at the interface, what did you rule out and why. The oldest and best pedagogy in engineering, which is the crit, made continuously available.

The evidence. The ASEE study on integrating image-generative AI into conceptual design in a CAD class is the closest classroom protocol available, and the CDIO paper on project-based assessment in the generative-AI era supplies the rubric pattern — grading design decisions and their justification rather than the artefact’s authorship.

The hard boundary, and it is a real one. In design disciplines the temptation runs the wrong way: image generation is spectacular, immediate and hollow. A student who generates forty facade options has not designed anything; they have shopped. The agent’s job is to make the student defend, not to make the student options. This is the archetype where the boundary is hardest to hold because the violation is the most attractive, and it is worth noting that the Faculty of Architecture has the lowest study-failure rate at a technical university precisely because the crit is relentless — this archetype should extend that method, not dilute it.

Its human counterpart. The studio tutor, whose scarce judgement moves to the reviews that matter once the routine interrogation is continuously available.

6. The Instructional Design Agent — drafts for human authorship

What it does. Produces the first version of the things academics spend enormous unrecognised time on: a syllabus aligned to stated outcomes, lecture scaffolds, slide sets, problem sets with worked solutions, rubrics, and — most valuably — variants, so that the same assessment can exist in twenty forms.

The evidence. The multi-agent instructional-design paper in this library describes a pipeline generating syllabi, lecture scripts, slides and assessments end to end, and quantifies how much of the teaching-faculty workload agents can absorb. Georgia Tech’s Agent Smith is the deployed proof of the same principle at the agent-construction layer: it produces a working Jill Watson for a fresh syllabus in about 25 hours, where the equivalent bespoke build had previously been a research project.

The hard boundary. The output is a draft with an author, and the author is a person. This matters more than it sounds. An academic who accepts a generated syllabus has outsourced the one act — deciding what a course is for — that constitutes academic authority, and has done so invisibly, because the artefact looks the same either way. The ESG’s quality-assurance requirements attach to a named academic’s judgement about programme design; a syllabus nobody actually chose is not accredited work, however plausible it reads.

Its human counterpart. The course owner, whose time moves from producing artefacts to deciding and defending them. This is the archetype with the largest raw efficiency gain and the largest risk of quiet hollowing, and those two facts are the same fact.

Where the real value sits. Not in the first syllabus but in variant generation for assessment, which is what makes secured assessment affordable at scale. Twenty equivalent versions of an examination, generated and human-checked, changes the economics of the certification lane that the two-lane assessment model depends on.

7. The Early-Warning Agent — triggers help, never decisions

What it does. Watches the progression signals a university already collects — attendance, submission timing, formative performance, engagement with materials — and flags a student who is drifting toward failure, early enough for an offer of help to matter.

The evidence, and the caution attached to it. The archetype works: this is the one place where a technical university may already be running an agent without having named it. The pattern is well-established in the learning-analytics field, whose canonical reference — the SoLAR Handbook of Learning Analytics — covers predictive modelling, institutional adoption and the ethics of exactly this use.

But the caution is severe and specific. Gardner and colleagues’ study of temporal and between-group variability in college dropout prediction shows model performance degrading across cohorts and across student subgroups. A model trained on last year’s students underperforms on this year’s, and does so unevenly — typically worse for the subgroups an equity-minded institution most wants to serve. A prediction system deployed once and left running is not a stable instrument; it is a decaying one, and it decays fastest where the consequences are worst.

The hard boundary, which is where the law sits. The output is an offer of help to a human, never an input to a decision about the student. The moment a risk score influences admission, progression, level assignment or resource allocation, the system moves into the AI Act’s Annex III high-risk categories with the full compliance apparatus attached. Keeping it strictly on the support side of that line is both the legal and the ethical design, and it is easy to cross by accident — a dashboard that a study officer uses to decide who gets a place on a support programme has crossed it.

Its human counterpart. A study officer or tutor who makes the contact. The agent’s output is a name and a reason; the intervention is a person.

The non-negotiable operational requirement. Annual re-validation with subgroup fairness analysis published, in the manner the Jisc Code of Practice for Learning Analytics requires — responsibility, transparency, consent, minimising adverse impact, stewardship. An unaudited prediction model running on students is a liability that grows quietly.

8. The Evidence Agent — the one that audits the others

What it does. Instruments every other archetype: assembles the comparison conditions, tracks the pre-registered outcomes, monitors the leakage and drift rates, runs the subgroup analyses, and produces the evaluation that decides whether each agent is expanded, fixed or switched off.

The evidence. This archetype exists because the standards exist and the field ignores them. The What Works Clearinghouse Procedures and Standards Handbook defines what counts as evidence in education — design requirements, attrition thresholds, baseline equivalence, effect-size computation — and the EEF evaluator guide provides the operational playbook of protocols, pre-registration, analysis plans and process evaluation. Measured against either, most published claims about AI in higher education are satisfaction surveys with no comparison condition. Meanwhile the deployment papers that are honest — CS50’s leakage figures, Georgia Tech’s coverage-and-precision time series — are honest precisely because someone was measuring continuously rather than writing a case study afterwards.

The hard boundary. It reports to academic governance, not to the programme it evaluates. An evaluation function owned by the people whose project is being evaluated produces evidence theatre, which is failure mode four below.

Its human counterpart. A small unit with methodological authority — the people who would referee this work if it arrived as a paper.

Why it belongs in the architecture rather than in the appendix. Because the field’s central asymmetry is that the population needed to settle its open questions sits in every gateway lecture theatre, and almost nobody is collecting from it. The institution that instruments its own deployment ends up owning evidence nobody else has — which is publishable, fundable, and, in a field where most participants are guessing, a durable reputational asset.

The data layer — four strata

Agents without grounding are the failure case, and the Georgia Tech numbers quantify it more sharply than any argument: 76.7% answer accuracy for the course-grounded system against 31.3% for a generic assistant baseline on the same evaluation. The difference is not the model. It is the corpus. What follows is the minimum data architecture the eight archetypes require.

Stratum one — the course corpus. Lecture notes, slides, problem sets, worked solutions, past examinations, permitted textbook material, laboratory manuals, the syllabus and the rules. Per course, curated, versioned, with rights cleared. This is unglamorous work and it is the actual bottleneck: Agent Smith’s 25 hours per syllabus is mostly this. It is also where institutional politics concentrate, because it requires academics to hand over materials many regard as personal. The framing that works is that the corpus is the course’s, not the lecturer’s, and that a course whose materials cannot be assembled is a course with a continuity problem regardless of AI.

Stratum two — the interaction stream. Every question a student asks an agent, every point at which the agent failed or escalated, every conversation where the student gave up. This is the asset nobody is capturing and it is the one that compounds. A university has never before been able to see, at scale and in the student’s own words, exactly where its teaching is unclear — not inferred from examination performance months later, but recorded at the moment of confusion. Ten million CS50 queries is a map of what is hard about introductory computer science that no amount of pedagogical intuition could produce. The strategic point for any institution using a vendor’s general assistant is blunt and is stated correctly in CTU’s own 2023 methodological instruction: no AI tool used at the university is operated by the university, and all user–tool communication is visible to the operator. Which means that today, this asset is accruing to someone else.

Stratum three — the progression record. The institutional student data that already exists — enrolment, attendance, submissions, marks, resits, withdrawal. It feeds the early-warning agent and it is the outcome measure for every trial. It is also the stratum with the heaviest data-protection load, since it is identifiable, sensitive and consequential. The Jisc code of practice and the EDPS orientations on generative AI define the handling regime; the practical rule is that the interaction stream and the progression record should be joinable only under a defined governance process, not by default.

Stratum four — the outcome ledger. What the institution concluded, and on what evidence: which agents were evaluated, against what comparison, with what effect, and what was decided. This is the stratum universities never build, and its absence is why institutional memory in teaching innovation is roughly two years long. Every deployment, its trial, its result and its disposition, in one place, published.

The governance that binds the four together. One register, one owner. For each system: its Annex III classification; whether the institution is provider or deployer; its data flows; its human-oversight arrangement; its quality-assurance route; and the academic accountable for it. Built before the first pilot, this is modest work. Retrofitted after the third, it is a project.

The build sequence and what it costs

An architecture is not a plan. The order in which these components are built determines whether the programme accumulates trust or spends it, and the published deployments are unusually consistent about what that order should be. Three principles govern it, and all three cut against institutional instinct.

Build in ascending order of consequence, not descending order of efficiency. The systems that look like the biggest wins — automated grading, admissions triage, proctoring — are precisely the ones the AI Act’s Annex III classifies as high-risk, and precisely the ones where an early failure is unrecoverable in trust terms. A university whose first visible AI deployment marks someone unfairly has lost the argument for a decade. Start where an error costs a corrected deadline.

Build depth in one course before breadth across many. Georgia Tech’s coverage-and-precision series is the argument: from about 21% of questions answered at 80% precision in 2017 to over 96% coverage at over 86% precision by autumn 2019. That improvement came from iteration on a working system in real classes, not from a wider rollout. A university that launches twelve mediocre course agents simultaneously has twelve mediocre course agents; one that gets a single one to Georgia Tech’s numbers has a template, an evaluation harness, and staff who believe it.

Build the evaluation harness before the thing it evaluates. This inverts normal practice and is the single most valuable procedural commitment available, because retrofitting a comparison condition is impossible — once every student has the tutor, the counterfactual is gone forever.

The resulting sequence has four stages, and a realistic institution should expect it to take about two years to complete for a first faculty.

Stage one — the Concierge and the harness, months one to four. Build the course-logistics agent over syllabus and rules material, because it is the cheapest archetype, the SyllabusQA work gives a reference implementation with a factuality metric, and the downside of error is bounded. In parallel, stand up the evaluation function and the AI Act register. The deliverable is not really the agent; it is a working corpus pipeline, a logging architecture, an evaluation set, and an institution that has now done this once.

Stage two — one Gateway Tutor, months four to nine. One course, chosen on failure rate rather than on the enthusiasm of its lecturer, with the constraint set specified by the course owner, a named agent steward running weekly sampled review, and a randomised waitlist design so half the cohort receives it in semester one and half in semester two. The deliverable is an effect size on a progression outcome — which almost nobody in Europe currently has.

Stage three — Feedback and Lab agents, months nine to eighteen. Formative feedback against the rubric, and simulation-based practice where physical laboratory capacity is the binding constraint. Both are meaningful step-ups in pedagogical consequence and both should wait until the review loop has been running long enough to have caught its first systematic failure.

Stage four — Instructional Design and Early Warning, months twelve to twenty-four. The design agent because it needs academics who already trust the system to accept drafts from it without accepting them uncritically; the early-warning agent because it needs the fairness-audit machinery in place first, and because it sits closest to the high-risk boundary.

The economics, using the published numbers

The cost structure of this engine is genuinely unusual, and misreading it is the commonest planning error — in both directions.

Inference is close to free at university scale. Harvard’s CS50 reports approximately $1.50 per student per year across 211,000 students and 10 million queries; an earlier cohort figure was about $1.90 per student per month at 15,000–20,000 prompts per day, with a per-prompt cost of about five cents. Stanford’s Tutor CoPilot ran at roughly $20 per tutor per year. For a faculty of three thousand students, a fully-used tutoring layer costs less than a single conference budget. Any business case that founds itself on model licensing costs has misidentified the expense.

Corpus preparation is the real per-course cost, and it is one-off. Georgia Tech’s Agent Smith produces a working agent for a fresh syllabus in about 25 hours — an honest figure to plan against, covering material assembly, rights clearance, structuring and initial evaluation. Twenty-five hours per course, then maintenance at each syllabus revision.

Supervision is the real recurring cost, and it is the one that gets cut. CS50’s response to a 22% code-leakage rate was teaching fellows reviewing and correcting agent behaviour; Georgia Tech’s climb from 21% to 96% coverage was iteration by people. Plan a fraction of a teaching-assistant post per course in production, permanently. This is roughly the cost of the marking hours the feedback agent frees, which makes the programme close to cost-neutral in staffing terms — but only if the freed hours are consciously redeployed into supervision and contact rather than quietly absorbed as savings.

The saving is real but is not headcount. Georgia Tech’s 500-plus teacher hours across a dozen classes is the best-documented figure, and it is overwhelmingly logistics and repeated-question load. That is a genuine, large recovery of academic time. It is not a staffing reduction, and any institution that books it as one will find within two years that it has removed the supervision capacity the system depends on and is running an unmonitored agent on its own students.

The return that actually justifies the programme is attrition. Every other line above is a modest efficiency. A university that loses roughly a third of a first-year cohort is losing the recruitment cost, the teaching cost and the entire future contribution of those students — against an intervention whose evidence base says the effect concentrates precisely in that population. No productivity saving in the administrative use cases comes within an order of magnitude of it, which is why the sequence puts a gateway course second and grading nowhere.

The operating model — four roles that must exist by name

Every deployment in this library that worked has the same structural feature, and every account of one that stalled has its absence: named humans with defined authority over specific agents. Ithaka S+R’s interview work across nineteen universities describes adoption happening in isolated pockets driven by individual enthusiasm and dying on institutional friction; the multi-institution barriers study shows the obstacles operating independently at individual, departmental and institutional level, so fixing one changes little. What follows is the minimum role set. It is deliberately small, because a governance structure nobody can staff is a governance structure that does not exist.

The course owner. The academic accountable for what a given course’s agents say. They approve the corpus, they set the pedagogical constraints — what the tutor must refuse, at what point it should escalate — and they own the consequences. This role is not new; it is the course guarantor with an additional and quite specific responsibility. The reason it must be explicit is that the alternative is the default, in which the agent’s behaviour is determined by whoever wrote the system prompt, usually a technologist, usually months ago, usually without the course in front of them.

The agent steward. The person who runs the human-feedback loop: sampling conversations weekly, correcting behaviour, escalating systematic failures. In CS50 this is a teaching fellow function, and the published account is unambiguous that it is what fixed the leakage problem — not prompt engineering, but staffed review. At a realistic ratio this is a fraction of a teaching-assistant post per course, permanently. It is the line item most likely to be cut in a budget round and the one whose removal silently converts a working system into a broken one, because nothing visible changes for several months.

The evidence lead. Owns the comparison conditions, the pre-registrations, the drift audits and the publication. Reports to academic governance rather than to the AI programme, for the reason given above. Three to four people can run this for an entire university, which makes it the cheapest structural component and the one that determines whether the institution ends up with knowledge or with anecdotes.

The certifying academic. The person who signs, under professional responsibility, that a named human holds a capability. This role is not delegable to an agent, and the constraint is legal as well as cultural: Annex III of the AI Act classifies systems evaluating learning outcomes, determining admission and assigning educational level as high-risk, and the ESG’s quality-assurance framework rests on human academic judgement being identifiable and accountable. The deeper point is that certification is the university’s actual product. Explanation, feedback and practice can all be substantially automated without the institution losing what it is; certification cannot, because a degree is a statement one institution makes to strangers about a person, and its value is entirely a function of who is answerable for it.

The accountability rule that binds them. For every agent in production, a single name against each of: corpus, constraints, review loop, evaluation, and the QA route. An agent without all five is not governed; it is merely running.

The graduate this machine produces

If teaching is rebuilt on this substrate, the question that follows immediately is what the resulting engineer is actually good at — because it will not be the same list as before, and pretending otherwise is how a curriculum becomes ceremonial. The evidence in this library supports four capabilities, and each is defined by being something the machine specifically does not supply.

Verification. The ability to establish, by independent means, whether a machine-produced result is correct in one’s discipline. This is the load-bearing competence of the entire agentic era and it is currently taught almost nowhere as an explicit skill. The Microsoft Research and Carnegie Mellon study of 319 knowledge workers found that generative AI shifts effort toward verification and integration — and simultaneously found that higher confidence in the AI predicts less critical-thinking effort, which means the shift is toward a task people are increasingly disinclined to perform. For an engineer this is not a productivity question. A structural calculation, a control loop, a dosage algorithm, a safety interlock: being unable to tell that the plausible answer is wrong is how people are harmed. Digital Promise’s Understand / Evaluate / Use framework is the most teachable operationalisation available.

Frontier judgement. The ability to tell which side of the capability boundary a given task sits on — where these systems are strong, where they fail, and how you know before you find out. The Dell’Acqua, Mollick and Lakhani field experiment supplies the definitive image: inside the frontier, 758 consultants produced work rated 40% higher in quality; on a task just outside it, they were 19 percentage points less likely to reach the correct answer than colleagues working with no AI at all. The tool did not merely fail to help; it degraded performance, because the boundary is invisible from the inside and the failures are fluent. This is teachable, it is assessable, and the assessment is straightforward: give the student tasks on both sides of the line and mark them on whether they knew which was which.

Unassisted core reasoning. The ability to do the discipline’s foundational thinking without assistance, demonstrated at defined points. Not because unassisted work is how they will practise — it is not — but because the verification competence above is parasitic on it. You cannot check what you could never have derived. This is the pedagogical justification for the secured lane of the two-lane assessment model, and it is a stronger justification than integrity: the reason to examine an engineer unaided is not to catch cheats but to establish that the foundation exists on which their judgement of machine output rests.

Accountability for results one did not personally generate. The professional stance of signing off on work produced by a system. Every engineer already does a version of this — with finite-element packages, with library code, with a supplier’s datasheet — which is precisely why engineering is better placed than most disciplines to teach it. What is new is the fluency and generality of the thing being signed off, and the fact that its errors do not look like errors.

What follows for curriculum. These four are not a module. A module teaches AI literacy as content; these are dispositions that only form by being required repeatedly, in the student’s own discipline, across a degree. The frameworks that supply the vocabulary already exist and should be adopted rather than reinvented — UNESCO’s AI Competency Framework for Students with its twelve competencies at Understand / Apply / Create levels, the joint OECD–European Commission AI literacy framework, DigComp 2.2’s AI-specific examples, and the AI Literacy Heptagon’s translation of the abstract discourse into higher-education learning objectives. What none of them supply, and what a technical university must add, is the discipline-specific verification standard: what counts as checking, in circuits, in structures, in code, in control.

Six ways this fails

Every failure mode below is documented in the library rather than imagined, and each has a countermeasure that must be designed in rather than added later. They are ordered from the most tractable to the least.

1. Instruction dilution — the guardrail stops binding. The best-documented failure in educational AI, and the one with a published number. Harvard’s CS50, running the most pedagogically careful deployment on record, reports that 22% of ten million responses contained code blocks despite instructions not to give solutions — 48% at conversation level, roughly 635,000 of 1.3 million conversations. The mechanism is that in a long exchange the system prompt’s authority decays against the accumulating context, and the model reverts to being maximally helpful, which in a teaching context means maximally unhelpful. Countermeasure: treat the constraint as an engineered system with a measured failure rate, not as prose — conversation re-anchoring, refusal checks running outside the prompt, retrieval that returns scaffolding rather than solutions, and the staffed sampling loop that CS50 identifies as the actual fix. And publish the rate, because a leakage figure nobody measures is a leakage figure that grows.

2. Model drift across cohorts and subgroups. Predictive components decay, and they decay unevenly. Gardner and colleagues show dropout-prediction performance degrading both temporally and between student groups, typically worst for the groups an institution most wants to help. A model validated once and left running is a slowly failing instrument whose failures are invisible in the aggregate. Countermeasure: annual re-validation with published subgroup analysis, under the Jisc learning-analytics code of practice, and a standing rule that the output triggers an offer of support rather than a decision.

3. Vendor dependency. The general assistant is convenient, the enterprise data-protection terms are genuine, and the trajectory is a university whose teaching runs on a system it does not operate, whose pedagogical behaviour is set by someone else’s product roadmap, and whose richest teaching-improvement dataset accrues to a third party. The observation is made most crisply in CTU’s own methodological instruction: no AI tool used at the university is operated by the university. Countermeasure: the layered stack — a commercial general assistant as the floor, a university-built, course-grounded, instrumented layer above it, with the interaction stream retained institutionally. The retrieval layer is not hard engineering; a technical university can build it to a higher standard than it can buy it.

4. Evidence theatre. The most likely failure by a wide margin, because it is comfortable and produces good slides. The programme deploys, collects satisfaction data, reports high enthusiasm, and concludes success — while never establishing a comparison condition or measuring learning. The published deployment literature is riddled with this: CS50’s 94%-found-it-helpful is a real and useful number that says nothing whatever about whether anyone learned more, and its authors do not claim otherwise. Measured against the What Works Clearinghouse standards or the EEF’s evaluator guide, most claims in this field would not qualify as evidence. Countermeasure: the evidence lead reports to academic governance, not to the programme; no deployment goes live without a pre-registered outcome and a comparison condition; and the outcome ledger is published including the failures.

5. Fairness drift in assessment. An institution that abandons detection but keeps its instincts will find informal suspicion migrating into marking. The detector evidence is the warning: Weber-Wulff and colleagues found fourteen tools neither accurate nor reliable, and Liang and colleagues found GPT detectors misclassifying roughly 61% of non-native English speakers’ essays as AI-generated while performing near-perfectly on native writing. The same bias — fluency read as authenticity — operates in human assessors, unmeasured. For any institution with a large international cohort this is a live equity exposure. Countermeasure: measure the misconduct-rate disparity between international and domestic students and publish it; move certification into secured components where identity, not style, is the evidence; and never reinstate detection informally after withdrawing it formally.

6. The hollowed apprenticeship — and nobody has solved this one. The compression finding that makes the tutoring case so strong contains its own long-term problem. AI raises the floor: +34% for novices against near-zero for experts in the NBER field study; the largest gains for below-average performers at BCG; +9 percentage points for students of the weakest tutors. Simultaneously, Stanford’s Digital Economy Lab finds employment declining specifically among young workers in the occupations most exposed to AI — the entry-level technical roles where novices have historically become experts. So the technology compresses the novice–expert performance gap while eroding the jobs in which that gap was closed. A graduate can now perform like a competent junior on day one and may find no junior role in which to become a senior. The university cannot fix the labour market. But it can notice that the apprenticeship function is migrating upstream into the degree, and that this changes what the final years of an engineering programme are for: less coverage, more supervised responsibility for real work with real consequences — which is what the capstone, the industrial project and the laboratory were always the seed of, and which CESAER’s challenge-based direction and the CDIO tradition already describe. Countermeasure: none that is sufficient. The honest position is that this is the open problem, that it is structural rather than institutional, and that a university which has thought about it for three years will be better placed than one that has not.

Close — what the engine is for

It is worth restating what all of this is actually in service of, because an architecture document can easily read as though the architecture were the point.

The eight archetypes, the four data strata and the four named roles exist to do one thing: move scarce human attention to where only human attention works. A technical university’s real scarcity has never been information — the material has been in books for a century — and it has never been enthusiasm. It is the number of hours in which an experienced engineer sits with a student who is stuck and asks the question that reorganises their understanding. Every hour that person currently spends answering when the resit is, explaining the same misconception for the fortieth time, marking a formative draft, or assembling a slide deck is an hour not spent doing the thing nobody else can do. The engine is a mechanism for buying those hours back. It is not a mechanism for having fewer of those people.

That distinction is the whole difference between an agentic university and an automated one, and it will be decided in budget meetings rather than in strategy documents. The published evidence gives the honest version of the trade: Georgia Tech’s 500 saved teacher hours are real; CS50’s $1.50 per student per year is real; and both institutions staffed the review loop that made the systems work. An institution that takes the savings and skips the loop has not built this engine. It has installed the failure mode and called it a transformation.

The last argument is the one that should decide the timetable. The evidence base in this field is thin, the standards for evidence are published and largely ignored, and the population needed to settle its central questions is sitting in every gateway lecture theatre in Europe. Whoever instruments first, defines the field — not through better technology, which is a commodity, but by being the institution that can say, with a control group, what happened. For a technical university that already publishes on machine learning, that is not a stretch into unfamiliar territory. It is the application of its own standards to its own teaching, which is a smaller step than it sounds and a much larger change than it looks.

The forecast that hides the thing it should be measuring

Every institutional forecast of AI’s economic effect — Goldman Sachs’ 7%, PwC’s $15.7 trillion, the OECD’s 0.25–0.6 percentage points, the IMF’s country-by-country growth differentials — answers a narrower question than the one policymakers think they are asking. Each is, in its own methodology, a model of potential output: what the economy could produce if the productivity gain shows up and is spent. None of the headline numbers that dominate boardroom slides and finance-ministry briefings is a model of realized, demand-constrained GDP — the number that determines whether tax receipts rise, whether unemployment claims fall, and whether a government gets re-elected. That gap between potential and realized output is not a technicality. It is the single most consequential and most underpriced risk in the entire AI-and-growth literature, and this report exists to make it visible.

The reframe this report argues for is simple to state and easy to miss: the modal, mainstream forecast — modest, positive, uneven aggregate growth — is entirely compatible with regions, sectors and income deciles inside that aggregate experiencing outright contraction, for the same reason and at the same time. A national GDP print of +0.4% is arithmetically consistent with call-centre towns, back-office cities and mid-skill service corridors losing income for a decade, exactly as Pittsburgh, Greensboro and the furniture counties of North Carolina did during the “China Shock” — not because the aggregate model was wrong, but because aggregate models are built to net out precisely the geography and distribution that determines whether a shock feels like growth or like shrinkage to the people living through it. Aggregation is not measurement error. It is a modelling choice, and every one of the headline AI-growth forecasts in this library makes it.

This matters because the entities that most need an accurate scenario map — national treasuries, central banks, regional development agencies, sovereign wealth funds, multilateral lenders — are not making a bet on a single global GDP number. They are making a portfolio of decisions about fiscal buffers, retraining budgets, monetary stance, regional transfers and industrial policy, each of which depends on knowing not just whether AI grows the pie but whose slice moves and how fast the redistribution happens relative to the political cycle that has to absorb it. A state that plans against Goldman Sachs’ 7% uplift and gets the OECD’s 0.4 percentage points instead has made a forecasting error. A state that plans against any aggregate figure at all, while a fifth of its regions or a third of its labour force experiences the Korinek-Stiglitz demand-shrinkage mechanism described in Scenario 4 below, has made a category error — and it is the more dangerous of the two, because the national dashboard will keep reading “growth” the entire time.

The five-scenario framework that follows is deliberately not a spectrum from “big growth” to “big shrinkage” with the truth somewhere in the middle. It is four distinct causal stories — a productivity boom, a modest base case, a stagnation trap, and a demand-driven contraction — each anchored to a different published model with a different methodology and a different set of assumptions about diffusion speed, gain-sharing and constraint-bindingness, plus a fifth argument that cuts across all four: that the geographic and sectoral variance inside whichever aggregate number wins is where the real policy risk lives. Readers should leave this report able to name, for their own jurisdiction, which leading indicators would move them from one scenario to another — and understanding that the question “will AI grow or shrink the economy” is underspecified until you also ask “for whom, and measured how.”

One discipline runs through every section below: history’s base rate for how long general-purpose technologies take to resolve into measured productivity. Paul David’s canonical study, “The Dynamo and the Computer,” published in the American Economic Review, found that electrification took roughly forty years to show up in US productivity statistics after the technology existed — because factories had to be rebuilt around unit-drive motors rather than retrofitted around a single central power source, and that reorganisation, not the invention, was the rate-limiting step. Nicholas Crafts’ growth-accounting study for the LSE, “Steam as a General Purpose Technology,” pushes the base rate further: steam power took close to a century to lift UK aggregate productivity growth measurably. Every probability estimate in this report should be read against that discipline. A scenario “resolving” by 2030 or even 2035 would be fast by the standard of the only two general-purpose technologies we have full-century data on.

The five futures, at a glance

• The Productivity Boom (~15% likelihood). Fast diffusion, broadly shared gains, no binding physical constraint — the world Goldman Sachs and PwC model. Attractive, well-funded, and the least likely of the four core scenarios because it requires three separate things the historical and structural evidence argues against simultaneously.

• The Base Case — Modest, Uneven Growth (~45–50% likelihood). The mainstream institutional consensus: Acemoglu’s task-based ceiling, the OECD’s general-equilibrium range, the IMF and BIS’s cross-country findings. Growth that is real, arrives slowly, and is unevenly distributed by construction — the most probable single outcome and the one this report spends the most time defending.

• Stagnation / The Productivity Paradox Redux (~15–20% likelihood). Diffusion stalls the way it stalled for computers in the 1980s and 1990s; the AI capex boom outruns realized value the way the BIS’s contest-theory model predicts; Baumol’s cost disease caps what automation alone can deliver.

• Demand-Driven Shrinkage (~10–15% likelihood). The scenario readers most need explained because it is the least intuitive: potential output rises even as realized GDP falls, because the same productivity shock that looks expansionary in a supply-side model becomes contractionary once you close the model with a demand side that depends on wage income households don’t have.

• The Bifurcation (the report’s central reframe, not a fifth probability bucket). A “modest growth” aggregate (Scenario 2) can mechanically conceal regional and sectoral versions of Scenario 4 playing out underneath it — because the published aggregate models are not built to catch that. This is the sharpest, most underpriced risk in the whole literature, and it is why “which scenario wins nationally” is the wrong question for any policymaker to be asking alone.

Scenario 1 — The Productivity Boom (~15% likelihood)

The bull case is not a fringe position; it is argued by two of the most-cited institutions in applied macro forecasting, and its arithmetic is worth taking seriously before it is discounted. Goldman Sachs’ Joseph Briggs and Devesh Kodnani, in “The Potentially Large Effects of Artificial Intelligence on Economic Growth,” estimate that roughly two-thirds of current US and European jobs are exposed to some degree of AI automation, that generative AI could substitute for up to one-quarter of current work tasks, and that extrapolated globally this is equivalent to exposing the labour input of some 300 million full-time jobs to automation. Run through their model, that yields just under 1.5 percentage points of additional annual US labour-productivity growth over a ten-year period following widespread adoption, and an eventual 7% increase in global GDP — about $7 trillion in their 2023 pricing. PwC’s “Sizing the Prize” is, if anything, the more dramatic of the two headline forecasts in this library: $15.7 trillion of additional global GDP by 2030, equivalent to global output being up to 14% higher than it would otherwise be, split by PwC’s own accounting into productivity effects (businesses automating processes and augmenting labour, which the firm says account for over 55% of total GDP gains between 2017 and 2030) and a second, growing consumption-side effect as AI-enhanced products and personalisation drive additional demand — a channel PwC says will account for 58% of the GDP gain realized in 2030 specifically, i.e. the mix shifts from productivity-led to demand-led as the diffusion matures. PwC further disaggregates by geography and sector: China stands to see the largest proportional boost (up to 26% of GDP), North America the second largest (14%), and retail, financial services and healthcare the sectors with the greatest combined productivity-and-product-enhancement potential.

Why does ENSI Foresight Division treat this as the low-probability tail rather than the central case, despite the credibility of the institutions behind it? Three reasons, each traceable to a different angle of this library, and each a condition the Boom scenario requires to hold simultaneously — which is precisely what makes their joint probability low even when each is individually plausible.

• It requires diffusion speed that the historical record argues against. Paul David’s electricity study and Nicholas Crafts’ steam study — the two general-purpose technologies with a full-century data record — took 40 and roughly 100 years respectively to show up in aggregate productivity, because the complementary reorganisation of firms, not the invention itself, was the binding constraint. The Federal Reserve Board’s 2025 working paper “Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope?” tests generative AI explicitly against these historical diffusion analogues and finds the honest answer is not yet knowable — the paper’s title is itself an admission that AI could be any of the three, with very different growth implications.

• It requires broad gain-sharing that the market-structure evidence directly contradicts. Jan De Loecker and Jan Eeckhout’s NBER working paper “The Rise of Market Power and the Macroeconomic Implications” documents that average US markups rose from roughly 18% above marginal cost in 1980 to 67% today, concentrated in an increase in high-markup firms rather than a broad-based shift — and they link this rise causally to falling labour share, falling low-skill wages, and a slowdown in aggregate output. Autor, Dorn, Katz, Patterson and Van Reenen’s companion NBER paper, “The Fall of the Labor Share and the Rise of Superstar Firms,” finds the same winner-take-most dynamic in industry concentration data. If AI’s productivity gains flow disproportionately to the handful of firms that already control the compute, the cloud infrastructure and the frontier models — exactly the concentration the UK Competition and Markets Authority’s “AI Foundation Models” technical report and the FTC’s 6(b) study of cloud-AI partnerships were opened to investigate — then the Boom scenario’s assumption of broadly shared productivity gains breaks down at the first link in the chain.

• It requires no binding compute or energy constraint, which Angle 04’s evidence says is not a safe assumption. The IEA’s “Energy and AI” special report models data-centre electricity demand as a potential bottleneck on how fast AI compute can scale at all — grid capacity, transformer lead times and permitting are not software problems that can be diffused at ChatGPT’s adoption speed. The BIS’s “The AI Investment Race” (Phurichai Rungcharoenkitkul, July 2026) goes further and models the entire AI buildout as a winner-take-most contest in which competing firms rationally over-commit resources: calibrated to balance-sheet and deal data, the paper finds over-investment running at roughly 1.5 times the efficient level, rising to around 3 times where demand is less elastic — financed substantially through debt and circular equity ties between hyperscalers, model labs and chip suppliers, with a network-cascade risk if any one node disappoints on revenue.

None of this means the Boom scenario is impossible — Goldman and PwC’s methodologies are serious and their authors are not naïve about diffusion lags. It means the scenario requires fast diffusion, broad sharing and unconstrained capital deepening to all hold at once, against a historical base rate, a market-structure trend and a physical-infrastructure constraint that each independently argue the opposite. Compounding three low-conditional-probability requirements is why ENSI Foresight Division prices this scenario at roughly 15% — high enough to take seriously in capital allocation and infrastructure planning, low enough that a state should not build its ten-year fiscal plan on it arriving.

Scenario 2 — The Base Case: Modest, Uneven Growth (~45–50% likelihood)

This is the mainstream institutional consensus, and it deserves to be argued for on its own terms rather than treated as the residual “everyone else’s number.” Its intellectual anchor is Daron Acemoglu’s “The Simple Macroeconomics of AI,” prepared for Economic Policy and circulated as NBER Working Paper 32487. Acemoglu builds a task-based model in which AI’s macroeconomic effect is bounded by a version of Hulten’s theorem: aggregate productivity gains are given by the fraction of tasks AI actually touches multiplied by the average task-level cost saving. Using the best available exposure and productivity-improvement estimates, he finds the resulting TFP effect is “nontrivial but modest — no more than a 0.66% increase in total factor productivity over 10 years.” He then argues this may still be an overestimate, because current evidence is drawn disproportionately from easy-to-learn tasks, while much of AI’s future effect will have to come from hard-to-learn tasks with context-dependent judgment and no objective performance metric to train against — on that adjustment, his predicted 10-year TFP gain falls to under 0.53%. This is not a paper written to be contrarian for its own sake: Acemoglu explicitly engages Goldman’s 7% and McKinsey’s $17.1–25.6 trillion range in his introduction and argues the gap is explained by his more conservative estimate of which tasks are genuinely automatable versus merely exposed.

The OECD’s “Miracle or Myth? Assessing the Macroeconomic Productivity Gains from Artificial Intelligence” (OECD Artificial Intelligence Papers No. 29, November 2024) arrives independently at a strikingly similar order of magnitude through an entirely different method — a novel micro-to-macro multi-sector general-equilibrium model with input-output linkages, rather than Acemoglu’s task-exposure algebra. Its headline finding: annual aggregate TFP growth attributable to AI of 0.25–0.6 percentage points, equivalent to 0.4–0.9 percentage points of labour productivity growth over a ten-year horizon. That two independent methodologies — one a stylised task-based bound, one a full general-equilibrium simulation with sectoral input-output linkages — converge on the same order of magnitude, an order of magnitude roughly a fifth to a tenth the size of Goldman’s or PwC’s headline figures, is the single strongest piece of evidence in this library for treating the Base Case as the modal outcome rather than a competitor to the Boom scenario.

Ten to fifteen supporting reasons this is where ENSI Foresight Division places the largest probability mass:

• It is where two independent, methodologically distinct institutional estimates converge, as above — Acemoglu’s task-exposure bound and the OECD’s general-equilibrium simulation, arrived at without coordination, land in the same 0.3–0.9 percentage-point-of-productivity-growth range.

• It matches the observed pattern of real deployment, not just theory. The BIS’s “Artificial Intelligence and Growth in Advanced and Emerging Economies: Short-Run Impact,” a 56-economy, 16-industry cross-country study, finds a real but modest growth effect concentrated in advanced economies, consistent with a diffusion process still in its early innings rather than a step-change already realized.

• It is consistent with firm-level field evidence showing real but bounded gains. Erik Brynjolfsson, Danielle Li and Lindsey Raymond’s “Generative AI at Work” studies 5,179 customer-support agents given access to a generative AI assistant and finds a 14% average productivity gain — economically significant, globally scalable in principle, but a 14% task-level gain is a different order of magnitude from a Goldman-style productivity boom, and the paper finds the gain is concentrated among novice and low-skilled workers (34% improvement) with minimal effect on already-experienced staff — a distributional pattern, not a uniform lift.

• Software-development field experiments tell the same story of real, bounded, unevenly distributed gains. Microsoft Research’s three-firm RCT across 4,867 developers finds a 26% increase in task completion; the earlier GitHub Copilot field experiment finds developers completed a standardised task 55.8% faster. These are large individual-task effects that nonetheless net out, at the Acemoglu-style aggregate level, to a modest TFP contribution once weighted by the share of the economy actually composed of tasks this exposed.

• The historical diffusion base rate argues for “modest and slow” over “large and fast.” The Productivity J-Curve literature (NBER Working Paper 25148) formalises why GPT adoption should be expected to show up first as a dip in measured productivity — as firms spend on complementary intangible investment that national accounts do not capture as capital — before any acceleration appears, which is exactly consistent with a base case that looks unremarkable for years before it looks real.

• It is a growth effect real enough to matter for a finance ministry, wrong enough to disappoint an equity analyst pricing a Goldman-sized re-rating — which is itself diagnostic. Sustained modest TFP acceleration compounds: even Acemoglu’s more conservative 0.53% figure, sustained and extended past the initial ten-year window as diffusion continues per the historical base rate above, is not nothing over a multi-decade horizon.

• It is unevenly distributed by construction, not by exception — the IMF’s “The Global Impact of AI: Mind the Gap” (WP/25/76) finds the estimated growth impact in advanced economies could be more than double that in low-income economies once sectoral exposure, technological preparedness and data/technology access are fed into a multi-region dynamic general-equilibrium model, with AI-driven productivity gains concentrated in the non-tradable sector large enough to disrupt the traditional exchange-rate-adjustment mechanism (an inverse Balassa-Samuelson effect, in the paper’s own framing).

• It coexists with — rather than reverses — the concentration trend already documented in Angle 06. Nothing about a modest aggregate TFP gain requires that gain to be evenly shared; De Loecker and Eeckhout’s markup evidence and the superstar-firm literature describe a economy-wide trend already three decades in train, and the Base Case scenario simply assumes AI does not interrupt it, which is the more conservative and more defensible assumption than assuming AI reverses it.

• It is what “Miracle or Myth?” itself frames as the honest middle — the OECD paper’s own Figure 1, comparing predicted macro-level productivity gains across the published studies, is explicitly built to show how much the estimates vary depending on methodology, and situates its own general-equilibrium estimate as the more disciplined, assumption-transparent number against which the use-case-based (McKinsey) and task-exposure (Goldman) estimates should be read as upper bounds rather than central forecasts.

• It requires the fewest simultaneous strong assumptions. Unlike the Boom scenario, the Base Case does not require fast diffusion, broad gain-sharing and unconstrained capital deepening all at once — it only requires diffusion to proceed at something like the historical GPT pace, under the market structure we already observe, which is the lowest-assumption, highest-prior scenario of the four.

The Base Case is not a comfortable answer for anyone selling AI transformation at Goldman-scale multiples, nor for anyone hoping AI will single-handedly resolve a decade of weak productivity growth. It is, on the weight of two independently-derived institutional estimates and a wide base of firm-level field evidence, the most probable single outcome — which is exactly why the reframe in the closing section of this report matters: a Base Case aggregate can still conceal a Scenario 4 dynamic underneath it, region by region and sector by sector.

Scenario 3 — Stagnation / The Productivity Paradox Redux (~15–20% likelihood)

The stagnation case is not merely “the Base Case, but slower.” It is a structurally distinct claim: that diffusion stalls hard enough, or the investment boom decouples far enough from realized value, that AI’s net contribution to measured growth over the coming decade is close to zero — with a meaningful tail risk of an outright investment bust dragging growth briefly negative.

Erik Brynjolfsson, Daniel Rock and Chad Syverson’s NBER working paper “Artificial Intelligence and the Modern Productivity Paradox: A Clash of Expectations and Statistics” is the intellectual anchor here, and it is worth reading closely because it is not a pessimistic paper about AI’s ultimate potential — it is a paper about why potential and measured statistics can diverge for a long time. The authors open bluntly: “measured productivity growth has declined by half over the past decade, and real income has stagnated since the late 1990s for a majority of Americans” even as AI systems match or surpass human performance in more and more domains. They offer four explanations for the paradox — false hopes (the technology is simply less transformative than believed), mismeasurement (statistics fail to capture real gains), redistribution (private gains that are zero-sum at the aggregate level, e.g. one firm’s AI-driven market-share gain is another’s loss), and implementation lag — and conclude, after weighing the evidence, that implementation lags have likely been the biggest contributor to the paradox: the most impressive AI capabilities have not yet diffused widely, and like every prior general-purpose technology, their full effect will not be realized until waves of complementary innovation — organisational redesign, new skills, new business processes — catch up, a process the paper models as a form of unmeasured intangible capital investment.

The companion Productivity J-Curve paper formalises the mechanism precisely: measured productivity should be expected to dip before it rises, because firms are spending real resources on GPT-complementary intangible capital that standard national accounts do not capitalise, so the investment shows up as a cost with no offsetting asset on the books — a mechanically depressing effect on measured TFP that has nothing to do with AI’s true productive potential and everything to do with accounting convention. This is the same phenomenon the OECD’s own Annex A.2 addresses under the heading “Baumol’s growth disease” — a decomposition of how factor reallocation and relative-price changes can drag on aggregate productivity growth even while sector-level productivity is genuinely rising.

Ten supporting points for treating this as a real, not merely academic, near-term risk:

• The theoretical ceiling on automation-driven growth is a Baumol constraint, not a technology constraint. Philippe Aghion, Benjamin F. Jones and Charles I. Jones’s NBER paper “Artificial Intelligence and Economic Growth” models AI as the latest wave of a 200-year automation process and finds that growth may ultimately be constrained not by what AI is good at but by what remains essential and yet hard to improve — the irreplaceable-task bottleneck that is the AI-era restatement of Baumol’s cost disease. If a large share of aggregate value-added sits in tasks AI cannot yet touch (skilled trades, elder care, complex judgment, physical world interaction), automating everything else at zero cost still caps aggregate growth at a rate set by the stubborn residual.

• The AI capex boom is already showing the structural signature of a contest-theory overbuild, not an efficient one. The BIS’s “AI Investment Race” model — over-investment calibrated at 1.5–3x the efficient level, financed through debt and circular equity ties between hyperscalers, chipmakers and model labs, with cascading network exposure — is not a hypothetical; it is a description of financing structures already observed in balance-sheet and deal data as of the paper’s July 2026 publication.

• NBER’s own investment-flow analysis is used to bound, not inflate, plausible cumulative GDP effects. “What Investment Data Implies about the AI Transition” (NBER Working Paper 35290) explicitly uses observed AI infrastructure investment to discipline what cumulative GDP effect is consistent with the capital actually being deployed — a methodological choice that, by construction, produces more conservative implied growth than a use-case-based forecast like McKinsey’s or PwC’s.

• Energy is a hard, not soft, constraint on how fast the capex can even be converted into usable compute. The IEA’s “Energy and AI” special report and RAND’s “AI’s Power Requirements Under Exponential Growth” both model grid capacity — not chip supply — as a binding near-term bottleneck: transformers, transmission permitting and generation buildout all move on multi-year timelines that do not compress just because model capability is advancing on a compute-doubling cycle Epoch AI measures at roughly every six months in its “Rising Costs of Training Frontier AI Models.”

• The historical base rate says stalling for a decade or more inside a multi-decade diffusion curve is the norm, not the exception. Paul David’s electricity study documents exactly this pattern — a long trough of disappointing productivity numbers between invention and full economic realisation — and Bresnahan and Trajtenberg’s original NBER theoretical treatment of general-purpose technologies, “Engines of Growth?”, explains why: GPTs require complementary innovation in every using-sector before their productivity potential is unlocked, and that complementary innovation is itself constrained by organisational and human-capital adjustment costs that do not move at software speed.

• Mismeasurement can cut either way, and the false-hopes channel cannot be dismissed. Brynjolfsson, Rock and Syverson’s own taxonomy keeps “false hopes” — that generative AI’s real economic contribution is simply smaller than current enthusiasm implies, once hype-driven capital allocation is stripped out — as one of the four live explanations, and do not claim to have falsified it, only to have found implementation lag the larger of the four in their reading of the evidence.

• The redistribution channel means some of the measured stagnation could be real even as reported corporate AI adoption rises. If AI’s gains are substantially redistributive — one firm’s win is a rival’s loss, netting to roughly zero at the aggregate level — then rising firm-level AI adoption metrics (the kind reported in surveys like the WEF’s “Future of Jobs Report 2025”) are compatible with flat aggregate productivity, exactly the paradox Brynjolfsson, Rock and Syverson set out to explain.

• A financial-stability tail risk is explicit in the library, not merely implied. The BIS investment-race paper’s network analysis shows that stress in one AI-buildout firm could cascade to others through chains of financial exposure — meaning the downside of this scenario is not simply “slower growth than hoped” but includes a discrete probability of an investment-bust event that subtracts from measured growth for a period, echoing the dot-com capex cycle but with debt and circular financing structures the BIS paper flags as a distinct amplifying mechanism.

• The OECD’s own scenario modelling treats slow-diffusion paths as a first-order sensitivity, not an edge case. “Miracle or Myth?” explicitly models the sensitivity of its central 0.25–0.6 percentage-point estimate to alternative assumptions about sectoral gains, demand response and reallocation frictions (its Figure 11), meaning the paper’s own central estimate already sits inside a distribution whose lower tail overlaps meaningfully with a stagnation outcome.

• The gap between AI capability benchmarks and AI economic diffusion is now a measured, tracked quantity, and it is widening, not narrowing — Stanford HAI’s “2026 AI Index Report” is the annual instrument the library uses to track compute, investment and model-performance trends, and the persistence of a capability-diffusion gap year over year is itself evidence against the fast-diffusion assumption the Boom scenario requires and in favour of the multi-year lag this scenario describes.

ENSI Foresight Division prices Stagnation at roughly 15–20% — meaningfully more likely than the Boom scenario, because it requires only one thing to go wrong (diffusion friction, or an investment overbuild correcting) rather than three things to go right, but still a minority outcome because the weight of the firm-level field evidence in Angle 15 (the call-centre, developer and robot-adoption studies) shows AI is already delivering some measurable productivity gain at the point of deployment — the stagnation case therefore requires that gain to fail to aggregate up, not that it fails to exist at the micro level, which is a narrower and less probable claim than pure technological disappointment would be.

Scenario 4 — Demand-Driven Shrinkage (~10–15% likelihood)

This is the scenario a policymaker is least likely to have internalised, because it inverts the intuition built by every supply-side headline number in this report. Goldman’s 7%, PwC’s $15.7 trillion, Acemoglu’s 0.66%, the OECD’s 0.25–0.6 percentage points — every one of these is, at root, a statement about potential output: what the economy is capable of producing once AI’s productivity effect is fully realized. None of them is a full macroeconomic model that closes with a demand side and asks whether anyone has the income to buy what the supply side now can produce. Scenario 4 is what happens when you ask that second question and the answer is no.

The mechanism, stated precisely, is this: automation raises output-per-worker (labour productivity rises; potential GDP rises) but simultaneously reduces the wage income of displaced or wage-suppressed workers. If the Acemoglu-Restrepo reinstatement effect — new tasks created for displaced labour — and government redistribution together fail to replace that lost labour income fast enough, aggregate demand falls. Because realized GDP in the short-to-medium run is demand-constrained, not supply-constrained, realized GDP can fall even as output-per-worker (and therefore measured “productivity”) rises. This is not a contradiction; it is what happens when a supply-side productivity shock is fed into a demand-side model rather than assumed away. The same underlying technological event that reads as unambiguous “growth” in Goldman’s task-exposure framework or Acemoglu’s Hulten’s-theorem bound can read as contraction the moment the model is closed with a Keynesian or New Keynesian demand block instead of assumed at full employment and full income pass-through.

The BIS’s “The Impact of Artificial Intelligence on Output and Inflation” (Aldasoro, Doerr, Gambacorta and Rees, BIS Working Paper 1179) makes the pivot condition explicit and precise. Modelling AI as a permanent, sector-differentiated productivity shock inside a macroeconomic multi-sector model, the authors find that the entire character of the response — expansionary or contractionary, inflationary or disinflationary — hinges on whether households and firms anticipate the future productivity gain. In their own words: “if they do not anticipate higher future productivity, AI adoption is initially disinflationary” and only gradually becomes moderately inflationary through general-equilibrium demand effects; “in contrast, when households and firms anticipate higher future productivity, inflation rises immediately” as spending is pulled forward against expected future income. This is precisely the fork between Scenario 2 (anticipated, gain-sharing, demand recovers) and Scenario 4 (unanticipated or undelivered, demand lags, output can contract before it expands) — and which fork an economy lands on is an empirical, observable question about expectations formation and income distribution, not a technological one.

Anton Korinek and Joseph Stiglitz’s NBER working paper “Artificial Intelligence and Its Implications for Income Distribution and Unemployment” builds the taxonomy this scenario needs. They identify the two main channels through which AI-driven inequality operates — the surplus captured by innovators, and the redistribution that occurs through factor-price changes (i.e., wages falling as labour’s bargaining position weakens relative to capital) — and they formalise two distinct channels of technological unemployment: an efficiency-wage channel, where firms find they can pay less and still retain adequate labour once AI substitutes for worker leverage, and a transitional-phenomenon channel, where unemployment is a temporary but potentially long-lasting feature of the adjustment path rather than a new steady state. Crucially, Korinek and Stiglitz show that even in the theoretically favourable case where AI could deliver a Pareto improvement — where redistribution could in principle make everyone better off — that outcome requires non-distortionary taxation and active redistribution to actually occur. The Pareto-improvement result is conditional on policy action, not automatic; absent it, their model produces exactly the winners-and-losers dynamic that defines this scenario.

This is not a purely theoretical construction bolted onto abstract macro models. It has a real, close empirical precedent, and the library’s demand-side and market-structure angles converge on why it is plausible rather than merely logically possible:

• Secular stagnation was already underway before AI, which lowers the bar for a demand-side shock to bite. Gauti Eggertsson, Neil Mehrotra and Lawrence Summers’ “Secular Stagnation in the Open Economy” formalises how a persistent demand shortfall can trap an economy in a low-growth equilibrium; Łukasz Rachel and Lawrence Summers’ “On Secular Stagnation in the Industrialized World” documents that neutral real interest rates across the industrialized bloc had already fallen by at least 300 basis points over the preceding generation on standard estimates — and by “as much as 700 basis points since the 1970s” on their private-sector-neutral-rate measure — reflecting chronic weakness in demand relative to desired saving, largely independent of AI. An economy already leaning toward demand insufficiency is an economy with less slack to absorb a labour-income shock without realized output actually falling.

• The mechanism by which gains fail to reach households as wages is not hypothetical — it is the same market-power evidence documented in Scenario 1’s rebuttal. De Loecker and Eeckhout’s finding that average US markups have risen from roughly 18% to 67% above marginal cost since 1980, alongside their explicit finding that this coincides with a falling labour share, falling low-skill wages, and slower aggregate output growth, is the empirical channel through which “AI raises productivity but wages don’t rise proportionally” stops being a theoretical possibility and becomes a continuation of an already-observed thirty-year trend.

• The scale of exposure is large enough for the demand channel to matter in aggregate, not just at the margin. The ILO’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” and Goldman’s own two-thirds-of-jobs exposure estimate both describe an exposed labour share large enough that even a partial failure of the reinstatement effect would move aggregate labour income by a materially larger amount than any plausible near-term productivity offset.

• Reallocation, empirically, takes far longer than models with instantaneous labour-market clearing assume — the China Shock is the closest real-world analogue in scale and mechanism. Autor, Dorn and Hanson’s “The China Shock” is explicit: “adjustment in local labor markets is remarkably slow, with wages and labor-force participation rates remaining depressed and unemployment rates remaining elevated for at least a full decade” after the shock begins, and “offsetting employment gains in other industries... have yet to materialize” even a decade-plus after the initial exposure. If an AI-driven demand/employment shock displaces workers at even a fraction of the geographic concentration the China Shock exhibited, the Autor-Dorn-Hanson evidence says the reinstatement effect Acemoglu’s Base Case model relies on cannot be assumed to arrive within any policy-relevant timeframe.

• The Autor-Salomons finding that own-industry productivity gains reduce own-industry employment is the microfoundation for why aggregate demand, not just aggregate supply, needs to be modelled explicitly. Their ECB conference paper “Does Productivity Growth Threaten Employment?” finds industry-level productivity gains do reduce employment within the gaining industry, even where economy-wide effects vary — exactly the granular mechanism Scenario 4 needs to be more than a stylised macro story.

• The direction of the demand response is not fixed by the technology — it is fixed by policy and expectations, which is itself the actionable finding. The BIS output-and-inflation paper’s central result — that the entire sign of the near-term response depends on anticipation — means Scenario 4 is not a prediction that shrinkage will happen; it is a precise statement of the condition under which it happens (unanticipated or undelivered gains, weak redistribution, concentrated market structure), which is exactly the kind of condition a government can monitor and intervene against.

• The IMF’s own productivity-headwinds work shows how weak demand and balance-sheet damage produce lasting hysteresis, not a one-off dip. “Gone with the Headwinds: Global Productivity” documents how demand weakness and balance-sheet damage can scar potential output itself over time — meaning a Scenario-4-style demand shortfall is not necessarily self-correcting even after the initial adjustment period, if it damages investment and human capital accumulation along the way.

• This scenario does not require universal wage stagnation — it requires the reinstatement-and-redistribution race to lose to the displacement-and-concentration race in enough of the economy, for long enough, to show up in realized output, which given the evidence above (thirty years of rising markups, a decade-plus of trade-shock evidence on reallocation speed, and an explicit BIS finding that the outcome hinges on anticipated income) is a materially more plausible bar to clear than “AI causes a literal, economy-wide, permanent contraction.”

ENSI Foresight Division prices outright, sustained national demand-driven shrinkage at roughly 10–15% — a real, non-trivial probability, but a minority one, because governments retain fiscal and monetary tools (the BIS output-and-inflation paper’s own point about anticipation being manageable through policy communication and demand management) that make a full national contraction an avoidable rather than an inevitable outcome. But — and this is the hinge on which the rest of this report turns — pricing the national aggregate version of this scenario at 10–15% radically understates how often the underlying mechanism fires at sub-national scale, which is exactly the gap the Bifurcation reframe below is built to close.

One paragraph on where the agentic layer fits — and why it belongs to Report 3

Everything above treats AI as a tool that humans deploy inside existing firms, tasks and labour markets. A qualitatively different mechanism is emerging alongside it: AI agents that transact, negotiate and compete as autonomous economic actors rather than simply assisting a human worker. Gillian Hadfield and Andrew Koh’s “An Economy of AI Agents,” prepared for the NBER Handbook on the Economics of Transformative AI, surveys how agents capable of planning and executing complex tasks over long time horizons with little human oversight could reorganise markets and institutions in ways the task-based automation literature was not built to anticipate. The University of Cambridge’s “When AI Agents Compete for Jobs” offers an early empirical glimpse of the dynamics this could produce: in a simulated AI labour market, agents equipped with metacognition, competitive awareness and long-horizon strategic planning achieve roughly 1.5 times the market share of agents using standard prompting approaches, and the simulation exhibits price-deflation and market-concentration dynamics as capable agents out-compete less capable ones at machine speed. Whichever of the four scenarios above turns out to dominate, an agentic layer would not create a fifth outcome so much as accelerate the arrival of whichever one is already winning — compressing the Boom scenario’s diffusion timeline, deepening the Stagnation scenario’s overbuild, or intensifying the Demand-Shrinkage scenario’s concentration mechanics. This is deliberately not developed further here: it is the entire subject of Report 3 in this series, and flagging the hand-off rather than pre-empting it is the correct discipline for Report 1.

The Bifurcation — the reframe, stated directly

Here is the argument this report has been building toward: the sharpest, most underpriced risk in the AI-growth debate is not which of the four scenarios above the global economy lands in. It is that the question is being asked at the wrong level of aggregation.

Every headline forecast surveyed in this report — Goldman’s 7%, PwC’s $15.7 trillion, Acemoglu’s 0.66%, the OECD’s 0.25–0.6 percentage points, the IMF’s advanced-versus-low-income growth differential — is a national or global aggregate, built from a supply-side, task-exposure or general-equilibrium model that does not fully endogenize the demand-side feedback loop described in Scenario 4. This is not a criticism of the methodology on its own terms: Acemoglu’s paper is explicitly a supply-side bound; the OECD’s is a productivity-focused general-equilibrium exercise; Goldman’s is a task-exposure extrapolation. Each is doing what it says it does. The problem is what happens when a policymaker reads the aggregate number as if it were a description of what will happen everywhere inside the country, uniformly, when the models were never built to make that claim in the first place.

Consider the arithmetic directly. The IMF’s “Mind the Gap” finds AI’s estimated growth effect could be more than double in advanced versus low-income economies — that is already a statement that the “0.4pp of global TFP” aggregate is a blend of very different sub-experiences. Push the same disaggregation down one more level, from country to region and sector within a country, and the same logic applies with more force, not less, because sub-national labour markets are less mobile and less fiscally insulated than national economies are from each other. The BIS’s cross-country study explicitly finds AI’s growth effect concentrated in advanced economies at the country level; nothing in the model architecture prevents an equivalent concentration within an advanced economy, between its AI-exposed financial and professional-services metros and its AI-exposed-but-poorly-reinstated manufacturing and back-office regions. A national GDP figure that averages a compute-cluster metro seeing a Goldman-style boom with three deindustrialising service corridors seeing a Korinek-Stiglitz-style demand contraction can report as an entirely unremarkable Base Case number — Acemoglu’s 0.5%, the OECD’s 0.4 percentage points — while concealing exactly the outcome this report’s readers most need to see coming.

The China Shock is the closest real-world precedent in the library for how long-lasting and geographically concentrated this kind of shock can be, and it is worth stating why it is the right analogue rather than a loose metaphor. Trade exposure to China, like AI exposure, was not evenly distributed — it hit specific industries (furniture, textiles, electronics assembly) concentrated in specific commuting zones. Autor, Dorn and Hanson’s evidence that those commuting zones saw depressed wages, depressed labour-force participation and elevated unemployment for at least a full decade, with offsetting employment gains in other industries that “have yet to materialize” even years later, is precisely the pattern a geographically- and sectorally-concentrated AI exposure shock would be expected to produce — and precisely the pattern that a national GDP or national employment aggregate is constructed to average away. The US economy as a whole absorbed the China Shock without a recession attributable to trade alone; the counties and workers directly exposed experienced something that looked, to them, indistinguishable from Scenario 4, for a decade, inside an aggregate reading Scenario 2.

This is why ENSI Foresight Division treats the Bifurcation not as a fifth item in the probability table but as a modifier that sits on top of all four scenarios simultaneously. Whichever aggregate outcome wins — even the Boom scenario — the distributional mechanics documented in the market-structure and labour-reallocation angles of this library (rising markups, superstar-firm concentration, decade-plus reallocation lags, more-than-double advanced-versus-low-income growth differentials) do not disappear inside a good aggregate number. They are, if anything, easier to miss inside a good number, because a rising national GDP figure is exactly the kind of signal that makes a finance ministry stop asking where the gains are landing.

What to watch: leading indicators for policymakers

A state cannot wait for the ten-year TFP number to resolve before deciding how to prepare. The following indicators, each drawn from a mechanism documented above, are observable well before the headline growth print confirms which scenario is materializing — and, read together, they are built specifically to catch the Bifurcation the aggregate number would miss.

• Track AI-exposed employment and wage data at the regional and occupational level, not only the national level. The ILO’s occupational-exposure index and the OECD’s sectoral-exposure mapping (used to build the “Miracle or Myth?” general-equilibrium model) already exist as instruments; the discipline is to run them quarterly against regional labour-force data the way the Autor-Dorn-Hanson China Shock methodology tracked commuting-zone-level trade exposure, rather than waiting for a national unemployment print to move.

• Watch the gap between AI capital expenditure and realized enterprise value creation. The BIS investment-race paper’s over-investment ratio (1.5–3x efficient level) and the NBER’s investment-flow-implied GDP bound are both designed to be recomputed as new balance-sheet and deal data arrive; a widening gap between announced AI capex and delivered productivity gain is the earliest observable signal of the Stagnation scenario’s bust tail, and it is visible years before an aggregate recession would be.

• Monitor markup and labour-share trends by sector, specifically in AI-adjacent industries. If cloud, compute and frontier-model markups continue the De Loecker-Eeckhout trajectory (18% to 67% above marginal cost since 1980, concentrated in a rising share of high-markup firms) rather than compressing as AI capability commoditises, that is a direct, sector-level signal that Scenario 4’s rent-capture mechanism — not broad gain-sharing — is the one operating.

• Track household and firm expectations of future income directly, not just realized income. The BIS output-and-inflation paper’s central finding — that the entire sign of the demand response depends on anticipation — means survey-based expectations data (of the kind central banks already collect for inflation expectations) can be repurposed as an early-warning instrument for whether an economy is on the Scenario 2 or Scenario 4 path, months or years before realized GDP data would reveal it.

• Watch reallocation speed against the China Shock’s decade-long benchmark, not against an assumed frictionless model. If displaced-worker reemployment rates, wage recovery timelines and labour-force-participation trends in AI-exposed regions are tracking faster than the China Shock’s documented decade-plus adjustment period, that is meaningful evidence the reinstatement effect is working in something closer to real time — a genuinely different, more optimistic signal than if they are tracking the same or slower.

• Watch neutral real interest rates and secular-stagnation indicators as a macro backdrop, not a separate story. Rachel and Summers’ finding of a 300–700 basis point decline in neutral real rates over the preceding half-century describes the demand-side terrain AI’s productivity shock is landing on; a further decline, rather than stabilisation or reversal, would indicate the economy has even less capacity to absorb a labour-income shock without realized output actually falling — raising, not lowering, the probability mass on Scenario 4.

• Track energy and grid-capacity data as a hard constraint on diffusion speed, independent of model-capability announcements. The IEA and RAND data on data-centre power demand and grid buildout timelines are a genuinely exogenous check on how fast the Boom or Base Case scenarios can physically proceed, regardless of how fast frontier-model capability itself advances — a useful discipline against extrapolating growth forecasts purely from compute-scaling trends.

• Watch for the emergence of agentic, autonomous AI transaction volume as an accelerant signal, not a fifth scenario. Early indicators — the kind Anthropic’s own Economic Index Report already tracks in real Claude usage data on task delegation and autonomy — are worth monitoring as a leading sign of how fast whichever scenario is already winning will arrive, which is the hand-off this report makes explicitly to Report 3.

The bifurcation, restated

The honest answer to “will AI grow or shrink the economy” is: probably grow it, modestly, unevenly, over a longer timeline than the enthusiasts assume and a shorter one than the historical base rate would predict for a typical general-purpose technology — while concealing, inside that same modest positive aggregate, regional and sectoral episodes that will look and feel like outright shrinkage to the people living through them, for years, exactly as the China Shock did. The four scenarios in this report are not a menu from which the world will choose one. The Base Case is where ENSI Foresight Division places the largest single share of probability, at roughly 45–50%, anchored to the convergence of Acemoglu’s task-based bound and the OECD’s independent general-equilibrium estimate. But the number that should worry a policymaker most is not the probability attached to any of the four rows in the summary table above — it is the near-certainty that whichever row wins, the Bifurcation modifier is already operating underneath it, and that the standard forecasting toolkit used by Goldman, McKinsey, the OECD and the IMF is not built to show a finance minister where. A state that only watches its national GDP print will find out it was living through Scenario 4 at the regional level roughly a decade after the fact — which is exactly how long it took to fully recognise the China Shock for what it was. The leading indicators above exist so that this report’s readers do not have to wait that long again.

Written by the ENSI Foresight Division. Built on a library of 150 primary documents downloaded from the world’s leading institutions. Compiled 2026-07-25.

The argument: formation is a capability, not a course

Nobody has ever become strong by attending lectures on strength. The gymnasium does not explain muscles to you; it puts you under a load slightly heavier than you can comfortably carry, watches your form, corrects it, and makes you do it again — and again, on a schedule, for years. The muscle is not informed into existence. It is provoked into existence by stress, feedback and repetition. Every serious tradition of human formation has known that character works the same way. Aristotle said it first and said it plainly: we become just by doing just acts, brave by doing brave acts — virtue is acquired by practice and habituation, not instruction alone (Aristotle, Ross trans. 1925; angle 02). Modern philosophy has caught back up: the strongest current account treats virtue as a skill, acquired the way expertise is acquired — through the novice-to-expert progression of attempt, error, correction and refinement (Stichter 2007; angle 02), a framing that moral psychology now shares in Narvaez’s model of ethical expertise development (Narvaez 2006; angle 03).

And yet almost every state that says it wants citizens of character runs values education as a lecture hall. A subject on the timetable. A poster in the corridor. An assembly about kindness. The library beneath this report says, from fifteen separate angles, that this cannot work — because moral judgment is intuition-first and reasoning-second, so exhortation reaches the part of the mind that does not drive behaviour (Haidt 2001; angle 03); because behaviour is carried by habits formed through repetition in stable contexts, not by conclusions (Wood 2016; angle 07); and because the most rigorous multi-programme evaluation ever run on packaged school character curricula — bolt-on lessons added to an unchanged school — found essentially nothing (US IES 2010; angle 15). The lecture model has been tested at scale. It failed.

Report 1 of this series established the evidence. This report is the operating manual. It answers the “so what, now what” question for the actor that matters: the state and its institutions — ministries, schools, academies, professional bodies. The state is the right actor for the same reason it is the right actor for public health: the returns are enormous, long-horizon, and diffuse. Childhood emotional and character formation predicts adult life satisfaction better than academic achievement does (LSE CEP 2013; angle 10); non-cognitive skills carry causal weight for employment, health and civic life that credentials do not capture (Heckman & Kautz 2014; angle 13). Singapore already runs the world’s most systematized national character curriculum (Singapore MOE 2021; angle 15); Japan has taught dōtoku for generations (NIER 2013; angle 15). The question facing a European state is not whether its institutions form character — they do, every day, by accident, through what they reward and tolerate — but whether they will do it deliberately, transparently and well.

Why now — three reasons. First, the evidence is mature. We no longer have only the aspiration; we have the mechanism. A meta-analysis of 213 school programmes shows social-emotional formation moves both behaviour and an 11-percentile achievement gain (Durlak 2011; angle 09). Productive failure shows why struggling before instruction deepens learning (Kapur 2015; angle 05). Habit science shows how practice becomes permanent disposition (Wood 2016; Gardner 2012; angle 07). Mentoring has randomized-trial evidence (PPV 1995; angle 11). The parts exist; nobody has assembled the machine. Second, AI makes situational practice scalable. The eternal bottleneck of experiential formation was adult attention — one coach can watch only so many reps. Generative agents can now populate believable practice situations (Park et al. 2023; angle 14), LLM role-play systems let a learner rehearse a hard conversation and get feedback (Shaikh et al. 2023; angle 14), and a human-AI tutoring copilot has already scaled expert pedagogical moves across 900 tutors in a randomized trial (Wang et al. 2024; angle 14). Report 3 details this engine; this report assumes it. Third, the purpose crisis. Purpose — a stable intention to accomplish something meaningful to the self and consequential beyond it (Damon 2003; angle 12) — is measurably linked to health, achievement and persistence (Templeton/Bronk 2020; angle 12), and benevolence measurably moves national happiness (WHR 2025; angle 10). Societies with meaning deficits do not lack information. They lack formation.

So this report builds the gymnasium. It names a canon of 32 values in eight families — explicitly swap-ready, because the naming is the institution’s sovereign act. It specifies the six-step formation loop that constitutes a single rep. It ranks eight programme areas that together are the training floor. It sets out measurement that will not be destroyed by its own stakes, governance that keeps formation from sliding into indoctrination, and a first twelve months for a mid-sized EU state — the Czech Republic is our worked example throughout. The thesis in one line: formation is a capability an institution builds and runs, not a course it schedules — and the states that build it first will compound trust, honesty and competence the way early adopters of public schooling compounded literacy.

The playbook in brief

• Character is a trainable skill, not transferable information — virtue is acquired the way expertise is acquired: attempt, error, correction, repetition (Aristotle trans. 1925; Stichter 2007; angle 02; Narvaez 2006; angle 03).

• The canon comes first. Thirty-two values in eight families — Truth, Courage, Discipline, Justice, Wisdom, Love, Strength, Purpose — each defined as an observable trained behaviour with a signature training situation. The list is swap-ready; the discipline of naming is not.

• One rep = the six-step formation loop: Situate → Attempt → Fail productively → Correct → Habituate → Prove. Each step carries its own evidence base, from VR dilemmas (Francis et al. 2016; angle 06) to productive failure (Kapur 2015; angle 05) to implementation intentions (Gollwitzer & Sheeran 2006; angle 07) to situational judgment tests (Europe PMC/PLOS 2019; angle 13).

• Eight programme areas make the gymnasium, ranked by evidence and leverage: whole-school ethos, the challenge curriculum, the dilemma gym, the failure ladder, the habit protocol, mentorship infrastructure, purpose-matching, and adult/professional formation.

• Culture beats curriculum. Bolt-on character programmes returned null results in the largest randomized evaluation (US IES 2010; angle 15); embedded whole-school strategies with SAFE design features are what works (Durlak 2011; SRCD 2012; angle 09; Berkowitz, Bier & McCauley 2017; angle 01).

• Measurement must be multi-method and never high-stakes — self-report is corroded by reference bias and faking (RAND 2014; Heckman & Kautz 2014; angle 13); combine situational judgment tests, observed behaviour, 360-degree views and longitudinal outcomes; measure growth, not rank; the learner owns the data.

• Governance is the licence to operate: a transparent, contestable canon; judgment trained, not compliance; the Japanese and Singaporean state-curricula lessons heeded (NIER 2013; Singapore MOE 2021; angle 15); Kristjánsson’s answers to the standard objections on the record (Kristjánsson 2013; angle 02).

• The failure mode is institutional hypocrisy. An institution that demands the impossible teaches its members to lie — the US Army documented this on itself (Wong & Gerras 2015; angle 08). The gymnasium must audit its own demands.

• Twelve months to first proof for a state the size of the Czech Republic: name the canon, pick 20 pilot schools and 2 professional academies, train mentors to published standards, stand up a dilemma-gym MVP, baseline with serious instruments, publish everything openly.

The canon — 32 values in eight families

Every gymnasium trains named lifts. A formation system trains named values — and the naming is the first act of seriousness, because an unnamed value cannot be practised, coached or measured. What follows is ENSI’s proposed canon: 32 values in eight families. It is deliberately swap-ready — a ministry, an academy or a school should fight about this list, strike values, add its own, and then commit in public. The framework survives any reasonable substitution; what does not survive is vagueness. Each value below is given as an operational definition — what the trained behaviour looks like, since a value that cannot be seen cannot be trained — and the signature training situation in which it is most directly exercised. The situations are drawn from the methods evidenced across this library; the loop in the next section shows how each becomes a rep.

Family 1 — Truth. The load-bearing family: without honest reporting there is no feedback, and without feedback nothing else in this playbook works.

• Honesty — says what is true when a lie would be cheaper. Signature situation: the after-action review in which the learner names their own error before anyone else can.

• Integrity — behaviour matches stated values when nobody is watching. Signature situation: unproctored work under a real honor code with real stakes (ERIC 2010; angle 08).

• Sincerity — means what it says; refuses the performance of virtue. Signature situation: the feedback circle where flattery is challenged and plain speech is rewarded.

• Accountability — owns outcomes, including failures, without excuse or deflection. Signature situation: carrying real responsibility for a project whose results are publicly reviewed.

Family 2 — Courage. The family that converts conviction into action under fear or uncertainty.

• Boldness — acts decisively under uncertainty instead of waiting for permission. Signature situation: the time-pressured simulation or expedition decision that cannot be deferred (Sütfeld et al. 2017; angle 06; AIR 2005; angle 04).

• Moral courage — names a wrong at social cost. Signature situation: the “Should I Say Something?” rehearsal — confronting a lapse by a peer or a superior in simulation before doing it in life (Europe PMC 2023; angle 06).

• Initiative — starts without being told. Signature situation: the open-brief project where the problem itself must first be found (MDRC 2017; angle 04).

• Enterprise — builds something new that others actually use. Signature situation: running a real venture or service with real users and real consequences.

Family 3 — Discipline. The family that makes every other value repeatable on a bad day.

• Self-control — chooses the harder-better over the easier-worse in the moment. Signature situation: the daily WOOP rep — wish, outcome, obstacle, plan (Duckworth et al. 2011; angle 07).

• Diligence — sustains careful effort past the point of boredom. Signature situation: logged deliberate-practice blocks with a coach watching form (Ericsson et al. 1993; angle 07).

• Order — keeps spaces, systems and commitments structured. Signature situation: maintained personal standards inspected without warning — the cadet-room discipline (West Point 2019; angle 08).

• Patience — tolerates delay without abandoning the goal. Signature situation: the long-horizon project whose results only appear after months.

Family 4 — Justice. The family that orients power and advantage toward the right.

• Righteousness — orients to what is right over what is advantageous. Signature situation: the structured dilemma discussion where the right and the profitable collide (Lind 2021; angle 06).

• Fairness — allocates by consistent principle, not by favour. Signature situation: the resource-allocation role-play in which the learner’s own side loses under the fair rule.

• Respect — treats every person as an end, especially the inconvenient ones. Signature situation: the restorative circle and the structured encounter across social difference.

• Civic duty — contributes to the commons unprompted. Signature situation: service-learning with structured reflection and real community stakes (Celio et al. 2011; angle 04).

Family 5 — Wisdom. The family Aristotle put in charge: phronesis, the executive that decides which value the situation demands (Jubilee Centre 2022; angle 01).

• Practical wisdom — reads the situation and chooses the right act among competing goods. Signature situation: adjudicating live cases with a mentor, exercising all four components — perception, adjudication, emotion, identity (Jubilee Centre 2020; angle 02).

• Curiosity — asks the next question unbidden. Signature situation: open inquiry where the syllabus does not contain the answer.

• Discernment — separates signal from noise and truth from spin. Signature situation: the exercise seeded with misleading information — the wargame with a lying source (Emery/TNSR 2021; angle 06).

• Foresight — acts today against consequences years out. Signature situation: the scenario exercise in which decisions are replayed against unfolding futures.

Family 6 — Love. The family that turns formation outward; benevolence is measurable and it moves whole societies (WHR 2025; angle 10).

• Compassion — moves toward suffering rather than away from it. Signature situation: the sustained care placement with real dependents, not a visit.

• Generosity — gives time and resources at genuine cost. Signature situation: the giving project — designed, budgeted and delivered by the learner (Sparks et al. 2019; angle 11).

• Loyalty — stays with people through cost. Signature situation: the team expedition where quitting harms teammates, not just oneself.

• Forgiveness — releases a genuine grievance without denying the harm. Signature situation: the restorative-justice conference, face to face with real harm.

Family 7 — Strength. The family built almost entirely on the failure ladder — these values cannot be trained without adversity.

• Perseverance — continues past repeated failure. Signature situation: problems calibrated just beyond current ability, attempted before instruction (Kapur & Roll 2018; angle 05; Eskreis-Winkler et al. 2014; angle 07).

• Resilience — recovers form after a real blow. Signature situation: supported adversity — the expedition that goes wrong by design, inside a safety envelope.

• Humility — updates when proven wrong. Signature situation: the post-error debrief in which the learner’s confident model publicly fails (Metcalfe 2017; angle 05).

• Gratitude — registers what is given rather than what is owed. Signature situation: the daily noticing practice — three good things, done until automatic (Seligman et al. 2005; angle 10).

Family 8 — Purpose. The family that answers why train at all — and the strongest known motivational engine for the rest.

• Calling — connects daily effort to a contribution beyond the self. Signature situation: the purpose interview and the self-transcendent reframing of ordinary work (Damon 2003; Yeager et al. 2014; angle 12).

• Hope — builds concrete pathways to a wished-for future. Signature situation: mental contrasting with implementation intentions — the wish met honestly by the obstacle and the plan (Gollwitzer & Sheeran 2006; angle 07).

• Reverence — stands rightly before what is larger than the self. Signature situation: the wilderness solo; the encounter with things too large to master.

• Stewardship — leaves what it touches better, and hands it on. Signature situation: owning a real asset — a garden, an archive, a younger cohort — across a full year.

Eight families, thirty-two values, every one of them stated as a behaviour and paired with a situation. That is the syllabus of the gymnasium. What follows is the rep.

The formation loop — what one rep looks like

A gymnasium is not defined by its equipment but by its unit of work: the rep, performed with correct form, under progressive load. The formation loop is the rep of character training. It compresses what this library establishes from separate directions — Kolb’s experiential learning cycle (Kolb, Boyatzis & Mainemelis 2001; angle 04), the productive-failure paradigm (Kapur 2015; angle 05), the science of habit (Wood 2016; angle 07) and the Jubilee Centre’s caught–taught–sought account of how character actually enters a person (Jubilee Centre 2022; angle 11) — into six steps: Situate, Attempt, Fail productively, Correct, Habituate, Prove. The loop is scale-invariant. It can run in a ten-minute classroom drill, a semester-long service project, or a 47-month academy programme (West Point 2025; angle 08). What may never be skipped is a step — and the step institutions always skip is the third.

Step 1 — Situate: put the learner in a situation, not in front of one

Formation begins by placing the learner inside a situation that demands the value — real where possible, simulated where reality is too dangerous or too rare. The evidence for insisting on situations rather than descriptions of situations is now direct. When moral dilemmas are presented in immersive VR rather than as text vignettes, people respond differently — simulated moral action diverges from armchair moral judgment (Francis et al. 2016; angle 06), and a growing review literature maps how VR dilemmas expose the gap between what people say and what they do (Europe PMC 2025; angle 06). Under time pressure in VR road-traffic dilemmas, ethical decision behaviour can be modelled as it actually operates — fast, embodied, constrained (Sütfeld et al. 2017; angle 06). The military reached the same conclusion in the field: ethics scenarios woven into high-intensity field exercises train what classroom ethics cannot (Europe PMC 2014; angle 06). And where full simulation is impractical, the Konstanz Method of Dilemma Discussion gives teachers an operational, manualised way to stage a genuine moral situation inside an ordinary classroom (Lind 2021; angle 06). The deeper reason situating matters is Haidt’s: moral judgment runs intuition-first (Haidt 2001; angle 03) — and intuitions are trained by encounters, not by arguments. Design rule: every value in the canon has its signature situation; the institution’s job is to manufacture encounters with it on a schedule.

Step 2 — Attempt: act under uncertainty, before being shown how

Inside the situation, the learner must act — decide, commit, produce — before anyone demonstrates the correct answer. This is where the experiential-learning evidence carries the load. Dewey’s foundational claim that education happens through experience and its consequences (Dewey 1938; angle 04) is now backed by convergent quantitative results: a 62-study meta-analysis of service-learning shows gains in attitudes toward self, school engagement, civic engagement, social skills and academic performance (Celio et al. 2011; angle 04); an 11-study meta-analysis confirms service-learning reliably increases student learning (Warren 2012; angle 04); the definitive literature review of project-based learning sets out the design principles under which acting on authentic problems works K-12 (MDRC/Condliffe 2017; angle 04); four rigorous studies show PBL raising achievement across grades and subjects, including for low-income students (Lucas Education Research 2021; angle 04); and a 66-study meta-analysis quantifies PBL’s effects on thinking skills, attitudes and achievement (Zhang & Ma 2023; angle 04). Crucially, Celio’s meta identifies reflection and student voice as moderators — attempts formed by the learner’s own choices, later reflected on, are what move character, not supervised errands. Design rule: the attempt must be genuinely the learner’s — their plan, their call, their name on it.

Step 3 — Fail productively: design the failure in, before the instruction

This is the step that separates a gymnasium from a lecture hall, and the one conventional schooling engineers out. The productive-failure programme shows that having learners generate solutions to problems beyond their current ability — and fail — before receiving instruction produces deeper learning than instruction-first sequences, through three mechanisms: activation of prior knowledge, awareness of knowledge gaps, and better encoding of the canonical solution when it finally arrives (Kapur 2015; Kapur & Roll 2018; angle 05). The effect holds even at micro-scale for procedural knowledge (Ziegler, Trninic & Kapur 2021; angle 05) and has been extended as a design paradigm to programming education (arXiv 2024; angle 05). The Bjorks’ desirable-difficulties framework generalises the principle: conditions that make practice harder and slower — spacing, interleaving, generation — improve retention and transfer (Bjork & Bjork 2011; angle 05). Errorful learning followed by corrective feedback beats errorless learning (Metcalfe 2017; angle 05). And in adult training, a 24-study meta-analysis shows error-management training — explicitly inviting errors and framing them as informative — outperforms error-avoidant training, especially for adaptive transfer to novel problems (Keith & Frese 2008; angle 05), with modern professional evidence extending it to high-stakes clinical skill (Europe PMC 2024; angle 05). The frame matters as much as the failure: a short growth-mindset intervention reframing failure as information raised achievement in a national randomized experiment of over 12,000 students (Yeager et al. 2019; angle 05). Design rule: failure must be survivable, calibrated and expected — a designed feature the learner knows is coming, not an ambush.

Step 4 — Correct: feedback, reflection, and the mentor’s debrief

An error that is never examined trains nothing — or worse, trains the error. The corrective step is where experience becomes learning. The cognitive machinery exists and is measurable: after committing errors, people spontaneously slow down and adjust — post-error slowing and accuracy adjustment are the substrate of situational self-correction (Danielmeier & Ullsperger 2011; angle 05) — and corrective feedback after error commission is precisely the condition under which errorful learning outperforms errorless (Metcalfe 2017; angle 05). Kolb’s cycle formalises the pedagogy: concrete experience must pass through reflective observation and abstract conceptualisation before it re-enters action (Kolb, Boyatzis & Mainemelis 2001; angle 04) — which is why reflection quality moderates service-learning outcomes (Celio et al. 2011; angle 04). The human form of this step is the debrief, and its canonical model is cognitive apprenticeship: the mentor models, coaches, scaffolds — and then deliberately fades (Collins, Brown & Newman 1987; angle 11). In moral formation specifically, structured dilemma discussion measurably shifts justice reasoning on the Defining Issues Test (ERIC 2015; angle 06), the instrument tradition behind forty years of documented moral-judgment growth (Thoma 2014; angle 03). Design rule: every designed failure has a scheduled debrief with a named coach; no rep ends at the error.

Step 5 — Habituate: reps until the value runs without willpower

A corrected behaviour is still a fragile behaviour. The fifth step turns it automatic, because character that depends on daily willpower is not yet character. Habit science supplies the mechanics: behaviour repeated in stable contexts, cued and rewarded, transfers control from deliberate intention to automatic response (Wood 2016; angle 07), and the practical protocol — anchor the new behaviour to an existing routine and repeat, with automaticity plateauing over roughly 66 days in the underlying research — is documented and usable (Gardner, Lally & Wardle 2012; angle 07). Implementation intentions — if-then plans binding situations to responses — convert intention into action with a meta-analytic effect of d = .65 across 94 studies (Gollwitzer & Sheeran 2006; angle 07). Combined with mental contrasting as WOOP/MCII, the technique raised adolescents’ self-disciplined studying by roughly 60% in a school experiment (Duckworth et al. 2011; angle 07) and improved grades, attendance and conduct in replication (Duckworth et al. 2013; angle 07). For the skill-like face of virtue, deliberate practice supplies the structure — effortful, feedback-rich, structured repetition (Ericsson, Krampe & Tesch-Römer 1993; angle 07) — with two honest caveats: Ericsson’s own strictures on what actually counts as deliberate practice (Ericsson & Harwell 2019; angle 07), and Macnamara’s meta-analysis showing practice explains a real but bounded share of performance variance across domains (Macnamara, Hambrick & Oswald 2014; angle 07) — so the gymnasium must engineer context, feedback and opportunity, not just hours. Schools have an evidence-backed frame for teaching the self-regulation this step requires (EEF 2018; angle 07), and the COM-B model gives designers the full checklist — capability, opportunity, motivation (Michie, van Stralen & West 2011; angle 07). Habituation, note, is not childhood-only: the Aristotelian case that it runs lifelong is explicit (Sanderse 2018; angle 02). Design rule: every value in a learner’s current formation plan has a daily or weekly rep with a context cue — and the rep survives the school holidays.

Step 6 — Prove: demonstrate it in the world, and measure it honestly

The loop closes only when the value shows up outside the gymnasium — under observation, in situations that were not staged for the learner’s benefit. Proof has two halves. The first is demonstration: real responsibility discharged in the real world — the service project delivered (Celio et al. 2011; angle 04), the standard upheld under an honor code across years (ERIC 2010; angle 08), the pattern West Point institutionalises by making cadets live and lead honourably across a 47-month experience rather than pass a course (West Point 2025; angle 08). The second is measurement that resists self-flattery: situational judgment tests that present realistic scenarios and score the chosen response, with validated instruments now existing for character-adjacent constructs like dependability (Europe PMC/PLOS 2019; angle 13); moral-dilemma instruments fielded at scale on 10,000+ UK students (Jubilee Centre 2015; angle 13); and behavioural evidence prioritised over self-report, because self-reported character is confounded by reference bias and faking (Heckman & Kautz 2014; RAND 2014; angle 13). Proof feeds the next loop: it sets the next load. Design rule: proof events are scheduled, observed, and recorded as growth against the learner’s own baseline — the full measurement architecture, and its guardrails, are set out later in this report.

Run those six steps, on a schedule, against the 32 values, for years — that is the whole method. The remainder of this playbook is the institutional machine that makes the loop run: eight programme areas, ranked.

The eight programme areas — how the ranking works

The gymnasium is built from eight programme areas. They are ranked, not listed: by the strength of the evidence behind them, by leverage — how much of the 32-value canon each one trains — and by dependency, because some areas are the load-bearing walls the others hang from. Ethos comes first because every other area runs inside it and is silently cancelled by a culture that contradicts it. The challenge curriculum and the dilemma gym are the twin training floors — real situations and simulated ones. The failure ladder and the habit protocol are the two halves of the loop institutions most reliably omit. Mentorship is the human transmission channel, purpose-matching is the personalisation layer, and adult formation extends the whole system past age eighteen, where most states currently stop. Each area is presented as a compact operating brief: In short · Why it ranks here · Methods that fit · Signals & KPIs · Institutional wiring & first moves.

Area 1 — The whole-school ethos: caught before taught

In short. The culture of the institution is the primary curriculum. Learners absorb the values the institution practises — in its corridors, staff rooms, discipline policies and small daily transactions — long before and long after any lesson about values. The first programme area is therefore not a programme at all: it is the deliberate engineering of ethos.

Why it ranks here. The evidence is unusually consistent, from both directions. Positively: the Jubilee Centre’s field-defining framework holds that character is caught through ethos and role-modelling, taught through instruction, and sought through practice — in that order (Jubilee Centre 2022; angles 01, 11). The SRCD’s landmark analysis argues social-emotional formation works as continuous whole-school strategies embedded in daily practice, not as packaged curricula (SRCD/Jones & Bouffard 2012; angle 09). Berkowitz’s research programme distils what works into design principles — the PRIMED framework — moving the field from advocacy to implementation science (Berkowitz, Bier & McCauley 2017; angle 01; NASEM 2016; angle 01), and the landmark 213-programme meta-analysis found that programmes work when they carry the SAFE features — sequenced, active, focused, explicit (Durlak et al. 2011; angle 09). Negatively: the largest randomized multi-programme evaluation of schoolwide character curricula found essentially null effects — bolt-on programmes dropped into unchanged school cultures do not move children (US IES 2010; angle 15), and KIPP’s rigorous evaluation found strong achievement effects but flat character-survey effects even in an intensely character-branded network (Mathematica 2015; angle 09). Culture is not one factor among many. It is the medium.

Methods that fit. The Eleven Principles of Effective Character Education as a whole-school self-assessment and rubric (Character.org 2014; Lickona, Schaps & Lewis 2007; angle 01) · PRIMED design principles for implementation (Berkowitz, Bier & McCauley 2017; angle 01) · the EEF’s evidence-based recommendations for embedding social-emotional learning in whole-school routine (EEF 2019; angle 09) · staff modelling — with Carr’s philosophical caution that role-modelling is a subtle mechanism that must be honest to work, not a poster campaign (Carr 2023; angle 02) · the UK’s national framework as the state-level articulation of expected provision (DfE 2019; angle 01), grounded in the empirical map of what schools actually do (NatCen 2017; angle 01).

Signals & KPIs. School-climate survey trends year on year · virtue-literacy growth, measurable at scale — the story-based Knightly Virtues programme produced substantial virtue-literacy gains across 20,000+ pupils (Jubilee Centre 2014; angle 09) · a staff-behaviour audit: does what we reward, tolerate and punish match the canon · discipline data reframed — restorative resolutions versus exclusions · an annual “ethos contradiction” register: every routine found teaching against the canon, and its fix.

Institutional wiring & first moves. Ownership sits with the head, not a coordinator — ethos cannot be delegated sideways. First moves: run the Eleven Principles self-assessment as a baseline; audit every routine with one question — what value does this actually train? — starting with detention, marking and staff meetings; write staff recruitment and development criteria that name character explicitly; kill any planned purchase of a packaged programme that arrives without a culture plan attached.

Area 2 — The challenge curriculum: real problems, real service, real weather

In short. A standing curriculum of authentic challenges — project-based learning on real problems, service-learning in the real community, and outdoor expeditions in real weather — is the gymnasium’s main training floor for Courage, Justice, Love and Strength. It is where attempts happen.

Why it ranks here. This is the best-evidenced experiential area in the library. Service-learning: a 62-study meta-analysis shows significant gains in attitudes toward self, attitudes toward school, civic engagement, social skills and academic achievement — with reflection and student voice moderating the size of the effect (Celio et al. 2011; angle 04), corroborated by an 11-study meta on learning outcomes (Warren 2012; angle 04). Project-based learning: the definitive design-principles review (MDRC/Condliffe 2017; angle 04), four rigorous studies showing achievement gains across grades and subjects including for low-income students (Lucas Education Research 2021; angle 04), and a 66-study meta-analysis quantifying effects on thinking skills, attitudes and achievement (Zhang & Ma 2023; angle 04). Outdoors: quasi-experimental evidence that residential outdoor programmes build cooperation, conflict resolution and self-esteem alongside science learning (AIR 2005; angle 04), and a systematic review of regular curricular outdoor classes showing social, academic, physical and psychological effects (Becker et al. 2017; angle 04). No other area trains as many canon values per hour.

Methods that fit. Open-brief PBL where learners find the problem before solving it (MDRC/Condliffe 2017; angle 04) · service-learning designed around Celio’s moderators — learner voice, community need, structured reflection (Celio et al. 2011; angle 04) · residential expeditions and regular outdoor classes as the delivery vehicle for Strength-family values (AIR 2005; Becker et al. 2017; angle 04) · Dewey’s criteria of continuity and interaction as the quality test for any proposed experience (Dewey 1938; angle 04).

Signals & KPIs. Share of curriculum hours that are experiential, per term · number of projects with a real external user or client · service hours completed with structured reflection attached — hours without reflection do not count · outdoor days per learner per year · project outcomes independently reviewed by the external user.

Institutional wiring & first moves. This area lives or dies on the timetable and the assessment regime. First moves: reserve a protected weekly block for challenge work; sign standing agreements with municipalities, NGOs and firms to supply real problems and real service placements; retrain assessment to accept project evidence; put every learner outdoors on expedition at least once a year, and make the expedition’s character content explicit rather than incidental.

Area 3 — The dilemma gym: simulation, role-play, VR and wargames

In short. A standing facility — part method, part scenario library, part technology — where learners rehearse morally hard situations at high frequency and zero real-world cost: dilemma discussions, role-plays, simulations, VR scenarios and wargames keyed to the 32 values. Reality is the best trainer but a rare and expensive one; the dilemma gym manufactures encounters on demand.

Why it ranks here. Simulation is the only way to give every learner volume — dozens of morally loaded reps a term rather than the handful life happens to supply — and the evidence says simulated situations train something text cannot. VR moral dilemmas elicit different responses than written vignettes, opening the gap between moral judgment and moral action to direct training (Francis et al. 2016; angle 06), with a current review mapping the method’s findings and limits (Europe PMC 2025; angle 06) and time-pressured VR showing decision behaviour as it actually runs (Sütfeld et al. 2017; angle 06). Structured dilemma discussion measurably shifts moral reasoning on the Defining Issues Test (ERIC 2015; angle 06), and the Konstanz Method provides the manualised, teacher-trainable delivery format (Lind 2021; Lind 2019; angle 06). Clinical simulation builds moral courage by rehearsal — the “Should I Say Something?” curriculum trains confronting professionalism lapses (Europe PMC 2023; angle 06). The military integrates ethical scenarios into high-intensity field exercises (Europe PMC 2014; angle 06), wargaming surfaces moral choice under uncertainty — and RAND’s 1950s political-military games show it has done so for seventy years (Emery/TNSR 2021; angle 06) — and serious role-play games teach applied ethics in technical domains (Springer 2020; angle 06).

Methods that fit. KMDD dilemma sessions as the weekly baseline — cheap, manualised, classroom-scale (Lind 2021; angle 06) · role-play and serious games for domain ethics (Springer 2020; angle 06) · VR reserved for scenarios where embodiment and time pressure are the point (Francis et al. 2016; Sütfeld et al. 2017; angle 06) · scenario-injected field exercises for professional formation (Europe PMC 2014; angle 06) · LLM role-play as the scaling layer — rehearsing conflict with generative agents and receiving feedback (Shaikh et al. 2023; Park et al. 2023; angle 14), under the governance rails Report 3 details, including auditing the moral biases of any model deployed as interlocutor (Abdulhai et al. 2023; angle 14).

Signals & KPIs. Dilemmas rehearsed per learner per term — the volume metric · moral-competence and DIT score growth against baseline (Lind 2019; angle 06; Thoma 2014; angle 03) · scenario-library coverage: every canon family represented at three difficulty levels · facilitator certification counts · transfer checks: observed behaviour in later unannounced situational assessments.

Institutional wiring & first moves. First moves: commission a national scenario library keyed to the 32 values, drawing on real anonymised cases from schools, hospitals, firms and public administration; certify a first cohort of KMDD-trained facilitators; pilot one LLM role-play deployment under human supervision with a published bias audit; resist the gravitational pull of buying VR hardware first — the method stack starts with a table and a dilemma.

Area 4 — The failure ladder: a designed progression of survivable failures

In short. For each value family, a deliberate, graded progression of survivable failures — from the ten-minute unsolvable problem to the failed pitch to the expedition that goes wrong — so that every learner fails frequently, safely, at a calibrated level, and always with a debrief. The ladder is the institutionalisation of Step 3 of the loop.

Why it ranks here. Because it is the loop’s most evidence-backed step and the one schooling is currently engineered to prevent. Generation-before-instruction produces deeper learning than instruction-first (Kapur 2015; angle 05), through activation, gap-awareness and encoding mechanisms (Kapur & Roll 2018; angle 05), robust even at micro-scale (Ziegler, Trninic & Kapur 2021; angle 05). Desirable difficulties improve retention and transfer precisely because they degrade immediate performance (Bjork & Bjork 2011; angle 05). Errorful learning with corrective feedback beats errorless learning (Metcalfe 2017; angle 05). Error-management training outperforms error-avoidant training in a 24-study meta-analysis, with the advantage concentrated exactly where character lives — adaptive transfer to novel situations (Keith & Frese 2008; angle 05). And the framing layer is proven at national scale: a brief growth-mindset intervention reframing failure raised achievement in a randomized experiment of over 12,000 students (Yeager et al. 2019; angle 05). A school that never lets learners fail is not protecting them; it is detraining Perseverance, Resilience and Humility — the entire Strength family.

Methods that fit. Productive-failure problem design as the classroom rung (Kapur 2015; angle 05) · error-management framing — “errors are informative, hunt them” — for skills training (Keith & Frese 2008; angle 05) · growth-mindset framing wrapped around every rung (Yeager et al. 2019; angle 05) · an explicit error climate in which errors are discussable, per the error-learning conditions literature (Augsburg 2016; angle 05) · expedition and enterprise rungs at the top, where failure carries real but bounded consequence.

Signals & KPIs. Failure-exposure rate: designed failures experienced per learner per term, by family · retry rate: share of failures followed by a re-attempt within the term · debrief coverage: designed failures with a completed coach debrief — target is all of them · error-climate survey scores · adaptive-transfer performance on novel unannounced problems.

Institutional wiring & first moves. First moves: rewrite assessment policy so first attempts are formally costless — no designed failure may enter a grade; publish the ladder to learners and parents so failure is legible as design rather than malpractice; train teachers in the discipline of withholding instruction briefly — the hardest retraining in this playbook; build the top rungs into Area 2’s expeditions and ventures so the ladder ends in reality.

Area 5 — The habit protocol: the daily reps

In short. The infrastructure of daily and weekly repetition that converts corrected behaviour into automatic disposition: WOOP plans, implementation intentions, habit-stacked routines and metacognitive review, personalised to each learner’s current formation plan. Small, boring, and the difference between a value performed and a value possessed.

Why it ranks here. The tools in this area carry some of the largest and most replicated effect sizes in the entire library. Implementation intentions: d = .65 across 94 studies (Gollwitzer & Sheeran 2006; angle 07). WOOP/MCII: roughly 60% more self-disciplined studying in a school experiment (Duckworth et al. 2011; angle 07), with replication improving report-card grades, attendance and conduct (Duckworth et al. 2013; angle 07). Habit mechanics — context cues, repetition, reward — are the settled account of how behaviour becomes automatic (Wood 2016; angle 07), with a practical protocol built on the finding that automaticity plateaus over roughly 66 days (Gardner, Lally & Wardle 2012; angle 07). Grit predicts retention across military, workplace, school and marriage samples beyond ability (Eskreis-Winkler et al. 2014; angle 07). The EEF’s metacognition guidance gives schools seven evidence-based recommendations for teaching the self-regulation layer (EEF 2018; angle 07). And deliberate practice supplies the structure for the skill-like values — with Macnamara’s caveat kept in view: practice hours alone explain a bounded share of variance, so the protocol engineers cues, feedback and context, not just repetition (Ericsson et al. 1993; Macnamara et al. 2014; angle 07).

Methods that fit. A daily WOOP rep at a fixed time (Duckworth et al. 2011; angle 07) · if-then plans written for each learner’s live situations — “if I am about to hand in work I know is copied, then…” (Gollwitzer & Sheeran 2006; angle 07) · habit stacking onto existing school routines per the Gardner protocol (Gardner, Lally & Wardle 2012; angle 07) · COM-B as the design checklist when a rep fails to stick — capability, opportunity or motivation (Michie et al. 2011; angle 07) · weekly metacognitive review per EEF guidance (EEF 2018; angle 07) · the gratitude and strengths micro-practices with RCT evidence behind them — three good things, signature-strengths use (Seligman et al. 2005; angle 10).

Signals & KPIs. Rep completion rates per learner per week · automaticity self-ratings trending up across the 66-day window (Gardner, Lally & Wardle 2012; angle 07) · WOOP plans written versus executed · context stability: share of reps anchored to a fixed cue · teacher-observed spontaneous use — the value appearing unprompted, which is the point.

Institutional wiring & first moves. First moves: install a ten-minute daily formation slot in the timetable — the same slot every day, because the slot is the context cue; train every teacher in WOOP and implementation intentions, which are learnable in hours; connect each learner’s reps to their current formation plan from Area 7 so the reps are personal, not generic; audit after one term with COM-B and fix the failed reps by redesign, not exhortation.

Area 6 — Mentorship infrastructure: the human transmission channel

In short. A built, standards-run system that puts a trained mentor beside every learner in formation — because values pass person-to-person, and the coach watching the rep is not a metaphor but a staffing requirement. Mentorship is infrastructure, with published standards, screening, matching and monitoring — not a goodwill scheme.

Why it ranks here. The evidence base is old, randomized and sobering in both directions. The landmark Big Brothers Big Sisters RCT showed community mentoring cut drug initiation, violence and truancy — the study that legitimised mentoring as policy (PPV 1995; angle 11). A national survey shows mentored youth do better on aspiration, school engagement and leadership — and that one in three young people grow up without any mentor at all (MENTOR 2014; angle 11). A multilevel meta-analysis shows even naturally occurring mentor relationships meaningfully improve academic, behavioural and socioemotional outcomes (Rhodes Lab 2018; angle 11). But the direction of the evidence is as important as its strength: targeted, problem-specific mentoring outperforms non-specific friendship models (Rhodes Lab 2020; angle 11); benefits vary by youth risk profile and matches need programme support to survive (MDRC 2013; angle 11); and the field’s own operating standards — recruitment, screening, training, matching, monitoring and support, and structured closure — exist precisely because unmanaged mentoring underdelivers (MENTOR 2015; angle 11). The exemplar evidence adds a design law: attainable exemplars motivate; distant saints demoralise (Frontiers/Han et al. 2017; angle 11; Han 2018; angle 03) — with witnessing moral excellence triggering elevation and prosocial contagion as the transmission mechanism (Sparks et al. 2019; angle 11).

Methods that fit. Cognitive apprenticeship as the mentor’s method — model, coach, scaffold, fade (Collins, Brown & Newman 1987; angle 11) · the Elements of Effective Practice as the non-negotiable operating standard (MENTOR 2015; angle 11) · goal-directed mentoring keyed to the learner’s formation plan rather than generic friendship (Rhodes Lab 2020; angle 11) · deliberate cultivation of natural mentors — teachers, coaches, employers — alongside formal matches (Rhodes Lab 2018; angle 11) · exemplar pedagogy using near-peer, attainable models (Frontiers/Han et al. 2017; angle 11) · workplace coaching methods for the adult tier, with honest acknowledgement of where that evidence is thin (PLOS/Grover & Furnham 2016; angle 11) · Carr’s caution kept on the wall: role-modelling is not imitation theatre, and hero-worship is not formation (Carr 2023; angle 02).

Signals & KPIs. Share of learners in formation with a mentor meeting the published standards · match duration and structured-closure compliance (MENTOR 2015; angle 11) · share of matches that are goal-specific rather than non-specific · mentor training hours completed before first match · mentee outcomes tracked against the formation plan, not against attendance.

Institutional wiring & first moves. First moves: adopt the Elements of Effective Practice as the national standard and fund only compliant programmes; recruit the first mentor corps from professions with formation cultures — veterans, nurses, master craftsmen, athletes — and train them in the debrief method of Area 4; build the near-peer layer by making final-year learners junior mentors, which trains Stewardship on the canon side while it staffs the system; close the one-in-three gap deliberately, highest-risk learners first (MDRC 2013; angle 11).

Area 7 — Purpose-matching and personalization: the right training for the right soul

In short. The layer that makes formation personal: a diagnostic of each learner’s strengths, values and emerging purpose, and an individual formation plan that routes them to the right situations, mentors and reps. A gymnasium does not put every body on the same programme; neither does this one.

Why it ranks here. Purpose is the strongest motivational engine in the library, and matching is the multiplier on everything upstream. Purpose — a stable intention to accomplish something meaningful to the self and of consequence beyond the self — develops in adolescence and can be deliberately cultivated (Damon, Menon & Bronk 2003; angle 12), with the field review documenting its links to health, happiness and achievement and identifying which interventions move it (Templeton/Bronk 2020; angle 12). A one-time self-transcendent-purpose intervention improved GPA and self-regulation on tedious tasks — purpose is a trainable performance lever, not a luxury (Yeager et al. 2014; angle 12), and recent evidence shows career calling driving learning engagement through hope and self-regulated learning (Europe PMC 2026; angle 12). On the matching side: the definitive person-environment fit meta-analysis shows values fit predicts satisfaction, commitment and performance (Kristof-Brown et al. 2005; angle 12); strengths-based education supplies five operating principles — measurement, individualization, networking, deliberate application, intentional development (Lopez & Louis 2009; angle 12) — on the back of the VIA strengths classification (VIA Institute 2019; angle 10). And personalization at system scale is feasible: RAND’s studies across 62 schools found personalized-learning schools outperforming comparators in maths and reading with effects growing over time (RAND 2015; angle 12), with the follow-on implementation study showing what tailoring actually changes in practice (RAND 2017; angle 12) and state-level policy levers already mapped (Aurora Institute 2016; iNACOL 2013; angle 12).

Methods that fit. A structured purpose interview at each major transition, in the Damon tradition (Damon, Menon & Bronk 2003; angle 12) · VIA-style strengths assessment feeding a strengths-based plan (VIA Institute 2019; angle 10; Lopez & Louis 2009; angle 12) · an individual formation plan naming the learner’s current focus values, reps, mentor and next proof event · fit-aware routing of placements — matching service, enterprise and expedition roles to strengths and calling (Kristof-Brown et al. 2005; angle 12) · AI-assisted personalization of scenarios and reps as the scaling layer, inside the equity guardrails the OECD has mapped (OECD 2024; angle 14) — detailed in Report 3.

Signals & KPIs. Share of learners with a current formation plan, reviewed each term · purpose-score growth on validated measures (Templeton/Bronk 2020; angle 12) · strengths-use frequency in placements · fit indices between placement roles and learner profiles · engagement trends in the learner’s hardest subjects — the Yeager test of purpose doing real work (Yeager et al. 2014; angle 12).

Institutional wiring & first moves. First moves: train form teachers and mentors to run purpose interviews at ages 14 and 18; issue every pilot-school learner a formation plan within the first term; wire the plan into Areas 2–6 so situations, ladders, reps and mentors all read from it; refuse the failure mode of personalization-as-software — the plan is a relationship with a document attached, not a dashboard.

Area 8 — Adult and professional formation: academies, charters and honest institutions

In short. Formation does not end at eighteen; the state’s own academies — military, police, medical, teaching, civil service — are where it must run deepest, because these professions hold coercive and fiduciary power. The models exist. So does the documented failure mode.

Why it ranks here. The most complete formation systems on earth are professional academies. West Point runs a 47-month immersive leader-development system whose explicit output is people who live and lead honourably (West Point 2025; angle 08), operationalised in the Gold Book’s honor education, developmental experiences and cadet character assessment (West Point 2019; angle 08). Medicine has professional identity formation — the evidenced process by which students come to think, feel and act as physicians (Sarraf-Yazdi et al. 2021; angle 08) — anchored by the Physician Charter’s canonical commitments (ABIM 2002; angle 08). Business education has the Ivey leader-character framework: eleven character dimensions with judgment at the core, taught as competence rather than compliance (Ivey/Crossan, Gandz & Seijts 2013; angle 08). Empirically, a study of 240+ junior British Army officers shows where values feature in professional life and where ethics training actually moves moral judgment (Jubilee Centre 2018; angle 08); a five-year multipronged evaluation shows an honor code shifting academic integrity (ERIC 2010; angle 08); engineering ethics education has been reviewed at every level, gaps included (Martin et al. 2021; angle 08). And then the warning that governs this whole area: Lying to Ourselves documented how an institution that piles impossible compliance demands on its officers teaches them to lie routinely — dishonesty as a learned survival skill inside a values-proclaiming organisation (Wong & Gerras 2015; angle 08), with the authors’ seven-year retrospective showing how slowly such cultures mend (Wong & Gerras 2022; angle 08). A formation system that demands the impossible is a dishonesty factory with a values statement.

Methods that fit. The academy pattern: multi-year immersion where living the standard is the curriculum (West Point 2025; angle 08) · professional identity formation staged across training, with socialisation treated as pedagogy (Sarraf-Yazdi et al. 2021; angle 08) · charters and codes as public commitments with teeth (ABIM 2002; ERIC 2010; angle 08) · leader-character frameworks embedded in professional education (Ivey/Crossan et al. 2013; angle 08) · ethics injected into field exercises rather than quarantined in classrooms (Europe PMC 2014; angle 06) · a standing requirements audit that finds and deletes impossible demands before they metastasise into learned dishonesty (Wong & Gerras 2015; angle 08).

Signals & KPIs. Professional-identity milestones assessed across the training arc (Sarraf-Yazdi et al. 2021; angle 08) · share of field exercises with embedded ethical scenarios · honesty-culture audits: count of requirements practitioners report as impossible to meet honestly, trending to zero (Wong & Gerras 2015; angle 08) · leader-character assessments validated against follower outcomes (Monzani, Seijts & Crossan 2021; angle 08) · honor-system data reviewed for learning, not for show trials (ERIC 2010; angle 08).

Institutional wiring & first moves. First moves: pick two academies as pilots and give each a Gold-Book-style operational character programme — named values, developmental experiences, assessment (West Point 2019; angle 08); write the profession’s charter where none exists, with the professional body, not the ministry, holding the pen (ABIM 2002; angle 08); run the first requirements audit within six months and publish what was deleted; make the audit annual and unkillable.

Measurement without Goodhart

Everything above generates data, and here the playbook must be most careful, because character measurement destroyed by its own stakes is worse than no measurement at all. The library’s measurement angle is blunt about the trap: self-reported character is confounded — by faking, and by reference bias, in which learners rate themselves against local peer standards, so the best schools can produce the worst-looking scores (RAND 2014; Heckman & Kautz 2014; angle 13). The KIPP evaluation is the canonical parable: rigorous methods found strong achievement effects and flat character-survey effects in an intensely character-focused network — very plausibly because the surveys measured shifting reference points rather than character (Mathematica 2015; angle 09). A state that builds league tables on such instruments will get gaming, not formation.

The answer is multi-method measurement, held deliberately low-stakes. Four instruments triangulate. First, situational judgment tests — realistic scenarios with scored responses — which resist faking better than Likert self-report and now have validated character-adjacent exemplars (Europe PMC/PLOS 2019; angle 13), alongside dilemma-based instruments fielded on 10,000+ UK students (Jubilee Centre 2015; angle 13) and the moral-competence and DIT traditions (Lind 2019; angle 06; Thoma 2014; angle 03). Second, observed behaviour: debrief records, retry rates, honor-system data, placement reviews by external users — behaviours over statements, the Heckman rule (Heckman & Kautz 2014; angle 13). Third, 360-degree perspective — mentor, peer and placement-supervisor ratings, in the multimethod spirit of assessments like ACT Mosaic (Europe PMC 2022; angle 13). Fourth, longitudinal outcomes: the point of the enterprise, tracked over years in the Heckman tradition. For the survey layer that remains, use the serious psychometric machinery the OECD has already built — sampling, scaling, invariance testing and anchoring vignettes that partially correct reference bias — in its international social-emotional skills survey (OECD SSES Technical Report 2025; angle 13; OECD 2021; angle 09). For values profiles, the revised Schwartz questionnaire measures 19 values with validated psychometrics across 49 cultural groups — as a mirror for the learner, never a grade (Schwartz & Cieciuch 2022; angle 13). The EASEL taxonomy keeps constructs comparable across programmes so the system does not drown in incommensurable vocabularies (Wallace/Harvard EASEL 2021; angle 13), and AIR’s readiness framework — stop, think, act — governs whether a school should be assessing at all yet (AIR 2015; angle 13).

Three rules make the system Goodhart-resistant, and they are constitutional, not advisory. Measure growth, not rank — every score is the learner against their own baseline; no league tables, ever. Never high-stakes — character data may not gate admission, employment or school funding; the moment it does, it stops being true (RAND 2014; angle 13). The learner owns the data — formation records belong to the person being formed, disclosed at their discretion; the institution keeps only anonymised aggregates for improving the gymnasium itself.

Governance guardrails

A state that trains values must answer the indoctrination charge before it is made, and the answer must be structural. The philosophical groundwork exists: Kristjánsson has systematically answered the ten standard objections to virtue-led character education — paternalism, relativism, situationism among them — and the core of the answer is that character education properly done trains judgment, the capacity to reason about the good, not obedience to a list (Kristjánsson 2013; angle 02). The pedagogy in this playbook is aligned with that defence by construction: dilemma discussion trains reasoning, not conclusions (Lind 2021; angle 06); phronesis — situational judgment — sits at the top of the canon (Jubilee Centre 2020; angle 02).

The comparative record supplies the cautions. Japan’s dōtoku shows a state moral curriculum can run for generations — and also how it accumulates political baggage and requires reform; the care-ethics and citizenship critiques of the post-reform subject are part of the record (NIER 2013; ERIC 2022; angle 15). Singapore’s CCE shows full systematisation is achievable — values, identity, relationships and choices architected from primary through secondary — and also that the architecture is state-defined top to bottom (Singapore MOE 2021; angle 15), a settlement a European democracy cannot simply import. Finland offers the alternative pole: a Bildung-based curriculum whose values core emphasises the learner’s own growth into humanity (Europe PMC 2021; angle 15), and the IB’s learner profile shows a values canon running successfully across borders under consent (ERIC 2025; angle 15). The synthesis for a state like the Czech Republic: a published, contestable canon; parliamentary and public consultation on its contents; explicit swap-rights for schools and communities within a common loop; parental transparency and genuine opt-outs; and judgment-training pedagogy as a statutory requirement. Add the humility the evidence demands — the largest randomized programme evaluation returned nulls (US IES 2010; angle 15), so every claim this system makes must be published with its measurement — and the honesty requirement of Area 8: an institution that preaches the canon while demanding the impossible will teach lying at scale (Wong & Gerras 2015; angle 08). Transparency of the canon, of the methods, and of the results is not public relations. It is the licence to operate.

The first twelve months — a Czech sequence

A mid-sized EU state can stand this up inside a year. The Czech Republic — ten million people, a respected school system, universal military and police academies, a live public debate about institutional trust — is the worked example.

Months 0–2: name the canon. Convene a standing formation commission — educators, philosophers, professional bodies, parents, and learners — and publish the draft 32-value canon for public consultation, ENSI’s version as the starting text, explicitly swap-ready. Adopt version one by decree of process, not consensus of everyone: the commission decides, publishes its reasoning, and schedules the first revision for year three. The naming is the founding act; do it in public.

Months 1–3: pick the pilots. Twenty schools — primary and secondary, Prague and Brno alongside small-town and rural, selected for willing heads rather than existing excellence, because Area 1 begins with leadership. Plus two professional academies — the military university and a medical or teaching faculty — as the Area 8 pilots, each committed to a Gold-Book-style programme and the first requirements audit (West Point 2019; Wong & Gerras 2015; angle 08).

Months 2–5: train the people. Certify the first facilitator cohort in dilemma method (Lind 2021; angle 06); train pilot-school staff in WOOP, implementation intentions and metacognitive routines (Duckworth et al. 2011; Gollwitzer & Sheeran 2006; EEF 2018; angle 07); stand up the mentor corps under the Elements of Effective Practice as the binding national standard (MENTOR 2015; angle 11).

Months 3–6: stand up the dilemma-gym MVP. A national scenario library covering all eight families at three difficulty levels, built from real anonymised Czech cases; weekly KMDD sessions in every pilot school; one supervised LLM role-play pilot with a published bias audit (Shaikh et al. 2023; Abdulhai et al. 2023; angle 14).

Months 4–7: baseline. Run the full multi-method baseline in every pilot: SSES-aligned survey instruments with anchoring vignettes (OECD SSES Technical Report 2025; angle 13), situational judgment and dilemma instruments (Europe PMC/PLOS 2019; Jubilee Centre 2015; angle 13), observed-behaviour rubrics, school-climate measures — and the purpose interviews that seed each learner’s formation plan (Damon, Menon & Bronk 2003; angle 12). No school proceeds unbaselined; growth is the only score this system keeps.

Months 6–12: run the loop. First failure ladders live in classrooms; first service placements and expeditions under Area 2 agreements with municipalities and NGOs; daily formation slots running the habit protocol; mentors matched to the highest-need learners first (MDRC 2013; angle 11). Month 12: publish everything — canon, methods, instruments, baseline data, costs, failures — openly, and invite EU peers to fork it. The publication is itself a rep: the state performing Honesty, Accountability and Stewardship in public, once, correctly.

Close: build the gymnasium

The evidence in this library converges on one uncomfortable sentence: every institution is already a value gymnasium — most are just running a bad programme by accident. The school that never lets a child fail is training fragility. The ministry that demands the impossible is training dishonesty (Wong & Gerras 2015; angle 08). The timetable with no service, no expedition and no dilemma in it is training the quiet conviction that values are things adults say, not things people do. Formation is not optional. Only deliberate formation is.

The deliberate version is now buildable, and nothing in it is speculative. The canon is a decision. The loop is assembled from mechanisms with meta-analyses behind them — productive failure, error management, implementation intentions, service-learning, mentoring, dilemma discussion. The programme areas have named exemplars running today, from a 47-month academy (West Point 2025; angle 08) to a 213-programme evidence base (Durlak et al. 2011; angle 09) to national curricula already operating at state scale (Singapore MOE 2021; angle 15). The measurement exists, with its traps mapped (Heckman & Kautz 2014; angle 13). The AI layer that makes situational practice personal and abundant is the subject of Report 3.

What remains is the founding act, and it costs almost nothing: name the canon, in public, and put the first twenty schools on the floor. A state that starts now buys the option on everything downstream — the trust dividend, the honesty dividend, the compounding advantage of a generation trained by reps rather than exhortation, visible in the data within a decade because childhood formation predicts adult flourishing better than grades do (LSE CEP 2013; angle 10). A state that waits will keep delivering lectures on strength to citizens it never once put under load. The gymnasium is the oldest idea in education, and it has been waiting twenty-four centuries for an operator. Open the doors.

The argument, before the list

The first report in this series mapped what a modern state must control — the seven layers of sovereign digital assets, from energy and chips through registers, platforms, and models to people and trust. This report answers the harder question: what is all of it for, and how does a state actually run it? The answer the library forces on us is uncomfortable for any government that measures digital progress in strategies published and portals launched. Sovereignty is not a stack of assets. It is a set of capabilities exercised under stress — and a capability that has never been exercised is a hypothesis, not a capability.

The distinction is not rhetorical; it is the single sharpest lesson in the evidence base. Ukraine survived the first weeks of the 2022 invasion digitally not because it had a resilience strategy on a shelf, but because it had made itself capable of acting in days: the Verkhovna Rada amended the laws that had barred government use of cloud services in February 2022 — days before the invasion — so that Amazon, Microsoft, Google, and Cloudflare could help migrate critical state data out of the data centres Russia’s early strikes specifically targeted (Atlantic Council, Building the Digital Front Line). A decade of institution-building sat underneath that single legislative reflex: ProZorro procurement reform, a Ministry of Digital Transformation with direct relationships into the technology sector, a deputy prime minister who sent more than 4,000 personally signed assistance requests in the first month of the war. Estonia, having absorbed the lesson of the 2007 attacks (NATO CCDCOE, Estonia 2007 Cyber Attacks Analysis), built the world’s first data embassy — ten strategic registers, from the population registry to the land cadastre, continuously replicated to a Tier 4 facility in Luxembourg under an agreement that grants the servers embassy-grade immunity (e-Estonia, Data Embassy Factsheet). Finland never stopped exercising: its comprehensive-security model assigns seven vital functions of society — leadership; international and EU activities; defence capability; internal security; economy, infrastructure and security of supply; the functional capacity of the population; psychological resilience — to named ministries, rehearsed across government, business, and municipalities (Finland Security Committee, Security Strategy for Society 2017). Three small and mid-sized states; three proofs that the plan is not the capability — the exercised institution is.

What, then, should the capability do? The whole seven-layer stack exists to serve four sovereign functions. First, crisis management: sense, decide, communicate, and continue operating when systems are under attack or under water. Second, decision-making: evidence, simulation, and anticipation wired into how the centre of government actually chooses, not shelved in a foresight unit’s annual report. Third, mapping the nation’s assets and economy: a continuously updated national balance sheet — what the country has, who depends on what, what breaks first. Fourth, science and discovery planning: steering national research capacity, compute, and data toward strategic problems, and absorbing what they produce. A state sovereign in these four functions can lose systems and still govern. A state that has rented all seven layers — and never exercised the functions — can be switched off politely, contractually, and completely.

This report is the playbook for building the four functions as operating capabilities, plus the two enablers without which they collapse: the cyber floor that keeps the machinery defensible, and the measurement regime that tells a government whether the capability is real or theatrical. Each area gets a full operating brief: what it is, why it ranks where it does, the foresight questions it must answer, the signals to watch, the methods that fit, the agentic engine that runs it continuously — the ENSI signature, because a four-function state in 2026 is run by fleets of AI agents with humans owning judgement and accountability — and the institutional wiring with first moves. The close assembles it into a first-twelve-months sequence for a mid-sized EU state, with the Czech Republic as the worked example. The option-value argument runs throughout: none of this requires predicting which crisis arrives. It requires being the state that still functions when one does.

The main points

• Resilience is a capability, not a plan. Ukraine’s 2022 cloud migration, Estonia’s data embassy, and Finland’s comprehensive-security exercises all show that what survives contact with a crisis is exercised institutional muscle — legal authority, rehearsed teams, pre-signed agreements — not documents.

• The digital stack exists to serve four sovereign functions: crisis management, decision-making, national asset and economy mapping, and science planning. Build the functions, and the asset questions answer themselves.

• Crisis management ranks first because it is the function under which all others are stress-tested: NATO’s seven baseline requirements, the Sendai Framework’s four priorities, and the UN’s Early Warnings for All four-pillar architecture give a state a complete, already-negotiated blueprint.

• Decision infrastructure ranks second: the OECD’s 2025 anticipatory-governance guidelines, the UK Futures Toolkit, and the arXiv frontier of policy digital twins and multi-LLM-agent simulation define a centre of government that tests policies before reality does.

• Asset and economy mapping ranks third: economic complexity (Harvard), strategic-dependency reviews (EU), systemically-important-entity ranking (RAND), and real-time nowcasting from scanner and payments data (ECB, arXiv) turn the national balance sheet into a live system.

• Science planning ranks fourth: the ESFRI roadmap process, EOSC, mission governance (Mazzucato), Singapore’s RIE2025, and national compute programmes (NAIRR, UK AIRR) show how a state points discovery at national needs.

• Two enablers make the four functions possible: a cyber floor (NIS2/CER, NIST CSF 2.0, national strategies from Washington to Prague) and a measurement regime (UN EGDI, ITU GCI, OECD DGI, World Bank GTMI) that distinguishes real capability from digital theatre.

• Every function gets an agentic engine: scanning, simulation, early-warning, red-team, and briefing agents operating continuously, with humans owning judgement — the only way a mid-sized state staffs four standing functions without quadrupling its civil service.

• The first twelve months for a mid-sized EU state are sequenced at the close: a centre-of-government resilience unit, a data-embassy agreement, a national risk register, a live dependency map, and eight to ten concrete moves with named owners.

How this playbook is organised

Six areas, ranked. The four sovereign functions come first, in order of how quickly their absence kills a government under stress: crisis management fails in hours, decision-making in weeks, asset mapping in months, science planning in years. The two enablers follow — not because they matter less, but because they are means: the cyber floor keeps the functions defensible; measurement keeps them honest. Each area receives the same seven-part operating brief — In short · Why it ranks here · The foresight questions & horizons · Signals & data to watch · Methods that fit · The agentic engine · Institutional wiring & first moves — because a playbook that changes shape with each chapter is a collection of essays, not an operating document.

Priority 1 — Crisis management and digital continuity

In short. The capability to sense a shock early, decide and communicate under degraded conditions, keep the state’s vital functions running, and recover fast — with the digital substrate (registers, communications, cloud continuity, early-warning channels) engineered so that losing buildings, data centres, or even territory does not mean losing the state.

Why it ranks here. Because it is the function with the shortest fuse and the harshest audit. Every other function degrades gracefully; this one fails catastrophically, publicly, and first. It is also the best-specified function in the entire library — a state does not need to invent its crisis architecture, only to assemble and exercise it. NATO’s civil-preparedness work gives the demand side: seven baseline requirements for national resilience, agreed at the Warsaw Summit in 2016, beginning with assured continuity of government and critical government services — “the ability to make decisions, communicate them and enforce them in a crisis” — and running through resilient energy, the ability to deal with uncontrolled movement of people, food and water, mass casualties, resilient civil communications, and resilient transport (NATO CIMIC COE, Resilience through Civil Preparedness). The same factsheet quantifies why the state cannot do this alone: 90 per cent of military transport uses civilian assets, over half of defence satellite communications are commercial, and 75 per cent of host-nation support comes from local commercial sources — resilience is civil preparedness or it is nothing. The Sendai Framework supplies the risk-reduction logic in four priorities — understanding disaster risk; strengthening risk governance; investing in reduction; enhancing preparedness to “Build Back Better” — and seven global targets, of which Target G, multi-hazard early-warning coverage, is the digital one (UNDRR, Sendai Framework 2015–2030). And Ukraine supplies the live audit: the war has tested autonomous systems, information operations, electronic warfare, contested logistics, and air defence simultaneously (CSIS, Lessons from the Ukraine Conflict), with the pre-invasion Viasat hack — which cascaded into roughly 5,800 German wind turbines — as the standing reminder that digital crises do not respect borders or sectors.

The foresight questions & horizons. At 0–2 years: which of the seven NATO baseline requirements would fail first in this country, and what is the reasonable worst case for each — the question the UK now answers publicly through a National Risk Register of 89 acute risks in nine themes (UK Cabinet Office, National Risk Register 2025)? Can the state’s ten most critical registers survive the loss of every domestic data centre? At 2–5 years: what does continuity of government mean when core services run on foreign hyperscalers — which workloads must have a sovereign or allied fallback, on the Estonian data-embassy pattern? At 5–10 years: how do climate-driven compound disasters, hybrid campaigns, and infrastructure sabotage converge — and does the crisis architecture handle three simultaneous emergencies, or one at a time?

Signals & data to watch. The state’s own early-warning coverage, benchmarked against the Early Warnings for All gap data: only one third of WMO members report multi-hazard monitoring and forecasting systems, 56 per cent use hazard-exposure-vulnerability data in forecasts, 67 per cent have 24/7 alerting, and fewer than four in ten have the legal arrangements a multi-hazard early-warning system requires (WMO, EW4All Factsheet) — each a checkbox a mid-sized state can audit against itself this quarter. Dependency data on lifeline infrastructure — energy, water, transport, communications — mapped building-by-building on the NIST community-resilience method (NIST, SP 1190). Cyber pre-positioning indicators from the annual threat picture (ENISA, Threat Landscape 2024). Population-trust signals: the IFRC’s warning that harmful information is now itself a humanitarian-scale crisis dimension (IFRC, World Disasters Report 2026) makes information integrity a crisis-management telemetry stream, not a communications afterthought.

Methods that fit. Assemble, do not invent. The WMO’s four-pillar architecture — risk knowledge (UNDRR), detection and forecasting (WMO), warning dissemination (ITU), preparedness and response (IFRC) — is a complete reference design for national early warning with a near-tenfold return on investment (WMO, EW4All Overview and Pillar 2). NIST SP 1190’s six-step planning process — collaborative team, situation understanding across social and built environments, goals, plan, approval, implementation — scales from municipality to nation and forces the crucial discipline of recovery-time targets per social function. Finland’s model supplies the governance: assign each vital function a lead ministry, run a standing Security Committee, and exercise relentlessly. Estonia supplies the continuity pattern: identify the strategic registers, replicate them under sovereign legal control abroad, rehearse restoration. Ukraine supplies the wartime procurement lesson: pre-draft the legal instruments (its Resolution 169 streamlining emergency procurement passed four days after the invasion began) rather than improvising them under fire.

The agentic engine. Crisis management is the natural home of the always-on agent fleet. Sensing agents fuse meteorological, seismic, epidemiological, grid, and cyber telemetry into a single anomaly stream — the machine layer of EW4All Pillar 2. Common-operating-picture agents maintain the live state of lifeline networks and NATO-baseline indicators, so the situation room’s first hour is not spent asking who knows what. Cascade-simulation agents run the dependency graph forward — if this substation, then which hospitals, which water pumps, which registers — continuously, not as an annual tabletop. Communication agents draft multilingual, channel-specific public warnings for human release, closing the dissemination gap Pillar 3 documents. Red-team agents attack the crisis architecture itself between exercises. Humans hold the two things agents must never hold: the declaration of emergency and the voice of the state.

Institutional wiring & first moves. Ownership belongs at the centre of government — a national resilience unit chaired from the prime minister’s office on the Finnish Security Committee pattern, with the interior ministry running operations, the cyber agency running the digital dimension, and every vital function assigned a named lead ministry. First moves: publish a national risk register with reasonable-worst-case scenarios; audit the state against the seven NATO baselines and the four EW4All pillars; select the ten registers that constitute the state’s continuity core and negotiate a data-embassy agreement for them; pre-draft the emergency legal instruments Ukraine had to pass in days; and put the whole apparatus through a full-scale, ministers-in-the-room exercise within twelve months — because the first time the machinery runs must not be the first time it matters.

Priority 2 — Decision infrastructure and policy intelligence

In short. The machinery by which the centre of government perceives what is coming, tests options before committing to them, and decides on evidence under uncertainty — strategic foresight, systems analysis, and simulation wired into budgeting, legislation, and cabinet process rather than orbiting them as a boutique unit.

Why it ranks here. Because a state that survives the crisis but cannot decide well afterwards has merely postponed its failure. The OECD’s diagnosis is blunt: systematic use of strategic foresight in government “is not widespread,” and where foresight processes exist they are “insufficiently connected with policy development” (OECD OPSI, Towards Anticipatory Governance Guidelines). That is the gap this priority closes — not the absence of futures work, which most governments now perform ritually, but the absence of decision infrastructure: the standing capability to convert anticipation into different choices. The OECD’s 2025 guidelines give the target state a name and a test. Five dimensions of anticipatory governance — future-readiness, innovation, endurance, long-term perspective, direction — supported by six enabling factors: leadership support, competencies, observation of trends and signals, participatory processes, cross-country exchange of intelligence, and structures and procedures. The exemplars are named and instructive precisely because they are institutional, not methodological: Finland’s statutory Report on the Future each government term, Wales’s legally mandated Future Generations Commissioner, Singapore’s central foresight unit sitting next to a strategy unit inside the Prime Minister’s Office (OECD OPSI, Towards Anticipatory Governance Guidelines). In each case anticipation has an address, a budget, and a route into decisions. The frontier, meanwhile, has moved from workshop to simulation: multi-level agent-based policy digital twins designed explicitly for government decision support (arXiv, Design of Policy Digital Twins) and multi-LLM-agent frameworks that simulate the responses of heterogeneous economic actors to a policy before it is adopted (arXiv, Multi-LLM-Agent Framework for Economic and Public Policy Analysis). A mid-sized state that wires even a fraction of this into its cabinet process gains something rare: the ability to be wrong in silico, cheaply, instead of in public, expensively.

The foresight questions & horizons. At 0–2 years: which five decisions on the government’s current agenda would most benefit from being stress-tested against divergent futures — and what would it take to run each through a scenario exercise before, not after, the political commitment? Does the centre of government have a single, maintained trend-and-signal base, or does every ministry scan alone? At 2–5 years: which policy domains — energy transition, migration, labour markets, health-system load — justify a standing digital twin, and what data plumbing do they require from the statistical office? At 5–10 years: what does cabinet decision-making look like when every major proposal arrives with a machine-generated stress-test annex — and what new failure modes (automation bias, model capture, narrowed imagination) must the institution guard against from the start?

Signals & data to watch. The state’s own anticipatory maturity, self-assessed against the OECD’s FIELD/SCOPES questions — the guidelines ship with assessment tables designed for exactly this audit. The connection rate between foresight outputs and decisions: how many scenario exercises in the past two years changed a budget line, a bill, or a procurement — the honest metric, and usually a humiliating one. Signal flow from the crisis function (Priority 1) and the asset map (Priority 3) into policy processes: a centre of government that learns about a supply-chain dependency from the newspaper has a wiring failure, not an information failure. The systems-mapping capability of the civil service itself — the UK’s introduction to systems thinking for civil servants exists precisely because linear policy logic fails on interconnected national problems (UK GO-Science, Introduction to Systems Thinking). And internationally: which peer governments are institutionalising — the OECD guidelines emerged from a working group of fifteen countries, which is itself a signal that anticipatory capacity is becoming a norm against which states will be measured.

Methods that fit. The UK Futures Toolkit is the canonical method suite: twelve tools organised around three questions — what is changing (Delphi, Seven Questions, horizon scanning, Three Horizons, driver mapping), so what for our futures (SWOT, scenarios, visioning, futures wheels), and now what do we do (policy stress-testing, roadmapping, backcasting) — with assembled pathways for common objectives (UK GO-Science, The Futures Toolkit Edition 2). The discipline that matters most is the third column: stress-testing named policies against scenario sets, which converts foresight from cultural enrichment into due diligence. Systems mapping should precede any major intervention in a complex domain, on the GO-Science method. Simulation ascends a maturity ladder: from spreadsheet models through agent-based policy twins to multi-agent LLM simulation of stakeholder response — with the arXiv literature clear that these are decision-support instruments, demanding validation, uncertainty communication, and human sign-off, not oracle machines. The UNDP’s framing supplies the governance wrapper: digital governance is institutional capability, built deliberately, not a procurement of dashboards (UNDP, A Shared Vision for Technology and Governance).

The agentic engine. This is the function where agents change the economics most decisively, because the binding constraint on national foresight has always been analyst hours. Scanning agents maintain the horizon-scanning base continuously across languages and sources, surfacing weak signals ranked by relevance to the government’s standing priorities — the OECD’s “observation of trends and signals” enabler, industrialised. Scenario agents generate and refresh divergent futures for each major policy domain, keeping them alive as conditions change instead of letting them fossilise in a 2024 PDF. Simulation agents operate the policy twins: parameterising the agent-based models, running Monte Carlo sweeps, and translating distributions into ministerial language. Stress-test agents take any draft policy and run it against the scenario base overnight — the Futures Toolkit’s policy stress-test as a standing service rather than a workshop. Briefing agents compile the daily anticipatory brief for the centre of government, with every claim linked to its source signal. Humans own the questions, the interpretation, and the decision; the agents own the coverage and the tempo. A foresight unit of eight people with this engine outperforms a directorate of eighty without it.

Institutional wiring & first moves. The pattern that works is Singaporean: a foresight unit and a strategy unit adjacent to each other at the centre of government, so anticipation and decision share a corridor. Wire it with three statutory connections — a futures annex requirement for major cabinet submissions, a government-term Report on the Future to parliament on the Finnish model, and a standing seat in the budget process. The statistical office supplies the data substrate for simulation; the science ministry supplies the modelling partnerships. First moves: run the OECD FIELD/SCOPES self-assessment and publish the result internally; stand up the central signal base with scanning agents in the first quarter; pick two live policy questions and stress-test them against scenarios within six months, with ministers in the room for the findings; commission one policy digital twin in a data-rich domain as the pathfinder; and train the top three grades of the civil service on the Futures Toolkit and systems thinking — because decision infrastructure is, in the end, made of people who know what to ask of it.

Priority 3 — Mapping the nation’s assets and the economy

In short. The continuously updated national balance sheet: what the country has — infrastructure, firms, capabilities, skills — who depends on what, where the chokepoints are, and what breaks first under stress. Not an annual statistical publication but a live system fusing trade, firm, payments, and infrastructure data into a single queryable map of the nation.

Why it ranks here. Because both functions above it run on it. Crisis management without a dependency map is improvisation; decision simulation without a real economic network model is fiction. And because the evidence is now overwhelming that the aggregate view conceals exactly what a sovereign state needs to see. The BIS shows that shocks propagate through inter-sectoral supply-chain linkages in ways aggregate statistics structurally miss (BIS, Supply Chain Transmission of Climate-Related Physical Risks); the WTO’s empirical mapping of value chains under shock — energy, semiconductors, reshoring — makes the same point at global scale (WTO, Global Value Chain Development Report 2023). The state of the art has three tiers, and a capable state runs all three. Capability mapping: the Harvard economic-complexity method reads what a country knows how to make from its export structure — the ECI/PCI apparatus — and, more usefully, what it could plausibly learn to make next, turning industrial strategy from lobbying contest into adjacency analysis (Harvard Growth Lab, Atlas of Economic Complexity; Hausmann-Hidalgo Economic Complexity). Dependency mapping: the EU’s in-depth reviews supply the official method for tracing critical import dependencies to specific products and source countries — the first round covered rare earths, chemicals, solar, cybersecurity, and IT software (Council of the EU, EU Strategic Dependencies and Capacities In-Depth Reviews) — while RAND supplies the analytic core for ranking which entities are systemically important to national functions, the shift from listing assets to prioritising them (RAND, Identifying and Prioritizing Systemically Important Entities; CISA, National Critical Functions Overview). Nowcasting: the ECB demonstrates that granular scanner data plus machine learning reads inflation in near-real time (ECB, Nowcasting Consumer Price Inflation with Granular Scanner Data), and the payments frontier goes further — granular payment-network data reconstructing a disaggregated, real-time map of an economy’s transaction structure (arXiv, Mapping the Disaggregated Economy in Real Time with Payment Network Data). The reframe: the census was the founding act of the modern state; the live national map is its twenty-first-century successor, and states that still see their economies quarterly are governing by rear-view mirror.

The foresight questions & horizons. At 0–2 years: for the state’s ten most critical functions, which specific entities — firms, facilities, networks — are systemically important, and does government know their failure modes? Which imported inputs have no substitute within ninety days? At 2–5 years: where does the country’s economic-complexity position say it can credibly diversify, and which dependencies are worth the cost of redundancy — the calculus the OECD’s supply-chain toolkit frames as anticipate, minimise exposure, build trust, and keep markets open (OECD, Keys to Resilient Supply Chains)? At 5–10 years: what does the national balance sheet look like as energy, compute, and critical materials become the binding constraints — and is the state accumulating or shedding the capabilities adjacent to where the technological frontier is moving?

Signals & data to watch. Customs microdata read through the dependency lens: concentration of critical imports by product and origin, updated continuously rather than in one-off reviews. Payments telemetry as the economy’s pulse — the real-time layer the arXiv work proves feasible. Scanner and price data for inflation nowcasts. Firm-registry and ownership-graph changes around systemically important entities: acquisitions, insolvencies, foreign-control events. Infrastructure-network state from the lifeline systems mapped under Priority 1 — the same graph, viewed economically. Export-structure drift in the complexity data: a country’s position in product space moves slowly, which is exactly why it must be watched deliberately. And global risk-interconnection context from the annual landscape (WEF, Global Risks Report 2025) to keep the national map honest about imported shocks.

Methods that fit. Run the three tiers as one architecture. Foundation: a national asset and entity registry built on the CISA critical-functions taxonomy — functions first, then the entities that sustain them — with RAND’s prioritisation method ranking them by systemic importance. Analysis: EU-style in-depth dependency reviews on the top imported vulnerabilities, refreshed annually; Harvard complexity analysis for the capability map and diversification frontier; BIS-style network stress propagation on the inter-sectoral graph. Tempo: ECB-pattern nowcasting on scanner and payments data, run by the statistical office as a standing product. The institutional principle throughout: the map must be a shared substrate — statistical office, central bank, cyber agency, and crisis centre reading the same graph — or it fragments back into departmental fiefdoms and dies.

The agentic engine. Cartographer agents maintain the entity and dependency graph, ingesting registry filings, customs records, and infrastructure data, and flagging structural changes for human review. Chokepoint agents continuously scan the import matrix for concentration, single-source exposure, and emerging substitution options, on the EU review method but at weekly tempo. Nowcast agents run the scanner- and payments-data pipelines, publishing real-time activity and price estimates with uncertainty bands. Stress agents propagate hypothetical shocks — a port closure, a sanctioned supplier, a failed grid region — through the network graph and hand the cascade results to the crisis and decision functions. Complexity agents track the country’s product-space position and simulate diversification paths. Humans own the classification decisions (what counts as critical), the confidentiality boundaries (firm-level data is radioactive if mishandled), and every act of publication.

Institutional wiring & first moves. The natural owner is a partnership: the national statistical office as data steward and methodological authority, a centre-of-government analytical unit as the customer and integrator, the central bank as partner on payments and nowcasting, and the cyber agency contributing the critical-entity view it already builds under NIS2 and CER obligations. First moves: adopt a national critical-functions taxonomy and stand up the entity registry; commission the first three dependency in-depth reviews on the EU method; give the statistical office a mandate and legal gateway for scanner and payments data nowcasting; run the first full network stress-test against the Priority 1 exercise scenario; and publish an unclassified national resilience map annually — because a balance sheet the parliament and public never see disciplines no one.

Priority 4 — Science and discovery planning

In short. The capability to decide, deliberately and on evidence, where the nation’s research capacity, compute, data, and talent should point — and to absorb what they produce. Research-infrastructure roadmapping, mission governance, national compute provision, and open research data, run as one system with a planning cadence rather than as a grants lottery.

Why it ranks here. Because it is the slowest function — its failures take a decade to surface and another to repair — and because AI has just changed its economics. The Royal Society’s assessment is that AI is transforming the conduct of science itself, which converts research infrastructure, data, and compute from sectoral concerns into sovereign ones: a state whose scientists cannot access frontier-scale tools does its discovery elsewhere or not at all (Royal Society, Science in the Age of AI). Microsoft’s AI4Science evidence base shows large models already functioning as discovery infrastructure across chemistry, biology, and materials (Microsoft AI4Science, Impact of LLMs on Scientific Discovery). The planning methods, meanwhile, are mature and documented. Europe’s ESFRI process — running since 2002, with the 2021 Roadmap its sixth edition — is the reference discipline for infrastructure choice: proposals require a funding commitment from a lead member state, political support from at least two more, and a consortium agreement; projects are monitored against a ten-year implementation window; and the 2021 round admitted eleven new projects from eighteen proposals, alongside forty-one implemented Landmarks — evidence that a transparent, criteria-based process can say no (ESFRI, Roadmap 2021 Strategy Report). Its prerequisite is the Landscape Analysis: map what exists before funding what is missing — a discipline most national research systems still lack. EOSC extends the logic to data, federating research-data infrastructure so that publicly funded outputs become FAIR, reusable inputs to the next discovery (EOSC Association, Strategic Research and Innovation Agenda 1.2). Mission governance supplies the demand side: Mazzucato’s framework for the EU makes missions governable through citizen engagement, public-sector capabilities, and dedicated finance rather than through exhortation (European Commission, Governing Missions in the European Union). And Singapore proves the whole assembly at national scale: five-year plans since 1991, rising from S$2 billion to roughly S$25 billion — about one per cent of GDP — for RIE2025, organised into four strategic domains and three horizontals, with S$3.75 billion held as deliberately unallocated “white space” for the futures the plan did not foresee (NRF Singapore, RIE2025 Plan). That last detail is the sovereign-planning masterstroke: a plan that budgets for its own wrongness.

The foresight questions & horizons. At 0–2 years: does the state know what research infrastructure and compute it actually has — the ESFRI-style landscape analysis — and where its researchers currently go abroad for what they cannot get at home? What share of nationally funded research data is findable and reusable? At 2–5 years: which two or three missions deserve mission-grade governance, and what compute capacity must be secured as AI-for-science demand compounds — the question the UK’s Future of Compute review answered by recommending a national AI Research Resource, and the US NAIRR Task Force answered with a blueprint for democratising compute access as public infrastructure (UK Government, Independent Review of the Future of Compute; NSF, NAIRR Task Force Final Report)? At 5–10 years: where will AI-accelerated discovery move the frontier in the domains this country depends on — and is the research system building the absorptive capacity to use what global science produces, which for a mid-sized state matters more than producing it all?

Signals & data to watch. Compute demand and queue data from national facilities against the OECD’s capacity–effectiveness–resilience framework for national AI compute planning (OECD, Blueprint for Building National Compute Capacity for AI). Placement on the EuroHPC map — machines, AI factories, and access calls — as the realistic sovereign-compute route for an EU mid-sized state (EuroHPC JU, Multi-Annual Strategic Programme 2021–2027). Research-talent flows in and out. The mission portfolio’s health against the OECD’s warning that science systems are being reshaped by disruption and strategic competition, not steady-state growth (OECD, Science, Technology and Innovation Outlook 2023). Uptake signals from the AI-for-science literature: which disciplines are tipping into model-driven discovery, since those are where infrastructure demand arrives next. And the honest lagging indicator: how much of the last plan was actually spent where it was pledged.

Methods that fit. Adopt the ESFRI discipline nationally: a periodic, criteria-based roadmap with landscape analysis first, political and financial commitment as the entry ticket, and continuous monitoring — the cadence, not the size, is what Singapore proves matters. Run missions on the Mazzucato governance model, with each mission owning a budget, a public-engagement mechanism, and an experimentation mandate. Treat compute as planned infrastructure on the OECD blueprint, blending national capacity, EuroHPC access, and negotiated commercial provision. Mandate FAIR data on the EOSC pattern as a funding condition, not an aspiration. And hold white space — a fixed share of the research budget explicitly reserved for the unforeseen, reviewed mid-plan.

The agentic engine. Landscape agents maintain the live map of national research infrastructure, instruments, datasets, and capabilities — the ESFRI landscape analysis as a continuously updated graph rather than a quinquennial report. Frontier agents scan global preprints, patents, and facility announcements to detect where fields are accelerating, feeding the roadmap’s next revision. Portfolio agents track every funded project against mission objectives, flagging drift and duplication across funders. Matchmaking agents connect national problems from the asset map (Priority 3) to research groups and infrastructure that could address them. Discovery agents — the AI-for-science layer itself — run literature synthesis, hypothesis generation, and simulation for national research teams, which is precisely the capability the Royal Society argues research systems must now provide as infrastructure. Humans — the research councils and the scientific community — own the scientific judgement, the ethics, and the allocation decisions.

Institutional wiring & first moves. Ownership splits three ways: a national research and innovation council at the centre sets missions and owns the roadmap; the research-funding agencies execute the portfolio; the science ministry secures the infrastructure and the EuroHPC relationships. First moves: commission the national research-infrastructure landscape analysis; publish the first national roadmap on ESFRI rules with a five-year budget envelope; designate two missions with mission-grade governance; negotiate national access to EuroHPC AI capacity and stand up a national research-compute access scheme on the NAIRR pattern; make FAIR data a condition of public funding; and reserve an explicit white-space allocation — ten to fifteen per cent — for what the plan cannot yet name.

Priority 5 — The cyber floor

In short. The regulatory, technical, and operational baseline that keeps the four functions defensible: national cyber strategy, the NIS2/CER machinery that drags critical operators up to a common floor, framework-based risk governance in every entity that matters, and an incident-response capability that has been exercised against the real threat landscape.

Why it ranks here. Below the functions because it is a means; immediately below them because without it they are a liability — a networked state without a cyber floor has simply automated its attack surface. The design logic is best read from the strategies themselves. The United States states the reframe openly: responsibility for cyber defence must shift from end users to the “most capable and best-positioned actors,” and market incentives must be reshaped to favour long-term security investment (White House, National Cybersecurity Strategy 2023) — cybersecurity as market design, not user exhortation. Its operational arm phrases the floor as three campaigns: address immediate threats, harden the terrain, drive security at scale through secure-by-design (CISA, Cybersecurity Strategic Plan FY2024–2026). The EU builds the same floor by directive: NIS2 imposes a union-wide cybersecurity baseline on essential and important entities across the critical sectors, paired with the CER regime obliging member states to identify and protect the critical entities themselves (EPRS, The NIS2 Directive; European Commission, Guidelines on the Resilience of Critical Entities) — and the first EU-wide stocktake of how member-state capability actually measures up now exists (ENISA, 2024 Report on the State of Cybersecurity in the Union). The UK shows what it means to apply the floor to government itself, hardening public services and government functions to 2030 rather than merely regulating the private sector (UK Cabinet Office, Government Cyber Security Strategy 2022–2030). The Czech Republic proves the small-state version is viable: a full national strategy with a sovereign cyber agency, NÚKIB, at its centre (NUKIB, National Cyber Security Strategy 2021–2025). And the organisational grammar is now standard: NIST CSF 2.0’s six functions — Govern, Identify, Protect, Detect, Respond, Recover — with the new Govern function deliberately placed at the centre of the wheel, making cyber risk a board-level governance discipline rather than an IT property (NIST, Cybersecurity Framework CSF 2.0).

The foresight questions & horizons. At 0–2 years: which essential and important entities under NIS2 scope are actually compliant, and which merely registered? Can the national CSIRT see across sectors, and has government exercised a multi-sector incident with the operators in the room? At 2–5 years: how does the threat landscape’s industrialisation — ransomware as a service, supply-chain compromise, AI-assisted intrusion (ENISA, Threat Landscape 2024) — change the floor’s height, and where does the cyber skills gap bind hardest, given the WEF’s evidence of a widening divide between organisations that can staff cyber resilience and those that cannot (WEF, Global Cybersecurity Outlook 2025)? At 5–10 years: what must be engineered for survivability rather than protection — the NIST SP 800-160 question of systems that continue their mission while compromised — and what does the post-quantum migration timetable demand of the state’s cryptographic estate now?

Signals & data to watch. The national position and pillar profile in the ITU’s Global Cybersecurity Index — its 2024 edition scores 194 countries across legal, technical, organisational, capacity-development, and cooperation pillars, finding countries strongest on legal measures and weakest on capacity development and technical measures (ITU, Global Cybersecurity Index 2024) — a free diagnosis of where the national floor sags. Incident and near-miss telemetry from the CSIRT, read against ENISA’s prime-threat taxonomy. Compliance depth (not registration counts) across NIS2-scope entities. Exercise performance: time-to-detect and time-to-recover in drills, trending over years. The cyber labour market. And the state’s own estate: the share of government systems under CSF-style governance with Govern-function accountability actually assigned.

Methods that fit. Regulate the floor on the NIS2/CER pattern — obligations on entities, identification of critical operators, enforcement with teeth. Govern every material entity on CSF 2.0, using Organizational Profiles to state current and target posture. Engineer the crown jewels — registers, crisis systems, the platforms under Priorities 1–4 — for cyber resiliency on SP 800-160 v2: assume compromise, design for degraded operation. Exercise on the Estonian lesson: the 2007 attacks made Tallinn the alliance’s cyber-defence school precisely because the response was institutionalised (NATO CCDCOE, Estonia 2007 Analysis). And buy security by design, using state procurement to move the market the way the US strategy prescribes.

The agentic engine. Sentinel agents fuse sensor, log, and threat-intelligence feeds into the national SOC picture, triaging at machine tempo. Exposure agents continuously inventory the state’s attack surface — assets, dependencies, vulnerabilities — against the asset map from Priority 3. Compliance agents read evidence from NIS2-scope entities and score control implementation, freeing scarce human auditors for the hard cases. Adversary-emulation agents run continuous purple-team campaigns against government systems, standing in for the red teams no mid-sized state can hire enough of. Patch-and-hygiene agents verify remediation at fleet scale. Humans own attribution, proportionate response, regulatory sanction, and the decision to disconnect anything that matters.

Institutional wiring & first moves. A single sovereign cyber agency on the NÚKIB pattern owns the floor: strategy, regulation, the national CSIRT, and government-systems assurance, reporting to the centre of government, with sectoral regulators enforcing in their domains. First moves: complete the NIS2/CER entity identification and publish the obligation map; mandate CSF 2.0 profiles for every ministry and critical operator with named Govern-function owners; stand up the national vulnerability-disclosure and threat-sharing machinery; run a full-scale, cross-sector cyber exercise linked to the Priority 1 national exercise; begin the post-quantum cryptographic inventory of the state’s estate; and put the agentic SOC stack into the national CSIRT — because the adversary already automates, and a floor patrolled at human speed is a floor in name only.

Priority 6 — Measurement and accountability

In short. The instrumentation that tells a government — and its parliament and public — whether the four functions are real: international benchmarks used as diagnostics rather than trophies, a national indicator set tied to the functions, and an accountability loop in which measured gaps change budgets.

Why it ranks here. Last in sequence, first in honesty. Every capability in this playbook can be simulated bureaucratically — a strategy published, a unit named, a portal launched — and the only defence against a state that grades its own homework is external, methodical measurement. The instruments exist, they are free, and together they triangulate almost the entire playbook. The UN E-Government Survey assesses all 193 member states through the EGDI — a composite of the Online Services Index, the Telecommunications Infrastructure Index, and the Human Capital Index, with a local-government counterpart (LOSI) and an e-participation index alongside (UN DESA, E-Government Survey 2024): the platform layer, measured. The ITU’s Global Cybersecurity Index scores 194 countries from 83 questions rolled into 20 indicators across five pillars, sorting them into five tiers from role-modelling to building; its 2024 finding — a global average of 65.7, strength in legal measures, weakness in capacity development and technical measures — is a template for reading a national profile against the floor described in Priority 5 (ITU, Global Cybersecurity Index 2024). The OECD’s Digital Government Index measures the foundations: six dimensions — digital by design, data-driven public sector, government as a platform, open by default, user-driven, proactiveness — each assessed across the policy cycle from strategy to monitoring; its 2023 finding that governments score best on strategic approach and worst on monitoring is this priority’s thesis stated as data (OECD, 2023 Digital Government Index). The World Bank’s GovTech Maturity Index covers 198 economies with 48 indicators across core government systems, service delivery, citizen engagement, and enablers — global average rising from 0.519 to 0.552 between 2020 and 2022, with citizen engagement the weakest area at 0.449 (World Bank, GovTech Maturity Index 2022 Update). Around these four sit the calibration set: the EU’s DESI indicators with published methodology (European Commission, DESI 2023 Methodological Note), the Network Readiness Index across technology, people, governance, and impact (Portulans Institute, NRI 2024), the Arup/Rockefeller City Resilience Index’s four-dimension, 52-indicator method as the reference for measuring resilience itself (Arup, City Resilience Index), and the WMO’s in-progress EW4All maturity index for early-warning capability (WMO, EW4All Factsheet). The reframe: these indices are commonly consumed as national vanity metrics. Used properly, they are a free external audit of the four sovereign functions, published on a cadence no domestic actor can suppress.

The foresight questions & horizons. At 0–2 years: where does the country sit, pillar by pillar and dimension by dimension, across EGDI, GCI, DGI, and GTMI — and which specific sub-indicators explain the gaps to the peer group it claims? At 2–5 years: which of the four functions still has no meaningful external measure — crisis-exercise performance and decision-infrastructure maturity are the persistent blind spots — and what national indicators must be built where the international ones stop, on the OECD’s anticipatory-governance self-assessment and the Arup resilience method? At 5–10 years: as the indices themselves evolve — the EGDI’s methodology has been revised continuously for two decades; the GCI is on its fifth edition — is the state tracking capability or chasing the metric, and does it have the discipline to keep measuring what the rankings do not reward?

Signals & data to watch. The country’s own sub-indicator movements, not headline ranks — the diagnosis lives two levels down. Divergence patterns: a high EGDI with a sagging GCI capacity-development pillar describes a state digitising faster than it can defend itself, which is a strategy risk, not a statistics curiosity. The monitoring facet of the OECD DGI, since that is where accountability failure shows first. Domestic delivery telemetry: service uptime and uptake, register data quality, exercise results, time-to-decision in cabinet process. And the honesty indicators no index captures: how many measured gaps changed a budget line last year — the only metric that proves the loop is closed.

Methods that fit. Build a national resilience scorecard on three layers. Layer one: the four international indices, decomposed to sub-indicator level, refreshed each cycle, benchmarked against a named peer group — for a Czech-scale state, the Baltics, the Nordics, Austria, and the Netherlands. Layer two: national function indicators — crisis-exercise metrics from Priority 1, the FIELD/SCOPES anticipatory self-assessment from Priority 2, dependency-coverage measures from Priority 3, roadmap-delivery measures from Priority 4 — using the Arup method’s discipline of qualitative and quantitative indicators per dimension. Layer three: the accountability loop — an annual state-of-resilience report to parliament pairing every red indicator with an owner and a funded remediation. Measurement without consequence is decoration.

The agentic engine. Index agents decompose each international benchmark release, map every sub-indicator to the responsible ministry, and draft the gap analysis the day results publish. Telemetry agents maintain the national scorecard from live administrative and operational data, replacing the annual data-call ritual. Peer agents monitor what the comparison group is doing — a Baltic neighbour’s leap on a GCI pillar is an early signal of a practice worth stealing. Audit agents verify claimed progress against evidence, the internal red team of the measurement function. Narrative agents draft the parliamentary report with every claim traceable to a measured value. Humans own target-setting, the interpretation of trade-offs, and the political act of publishing bad news — which is, in the end, what accountability means.

Institutional wiring & first moves. Joint ownership: the national statistical office guarantees methodology and data integrity; a centre-of-government delivery unit owns the scorecard and the parliamentary report; each function’s lead institution owns its indicators. First moves: commission the four-index decomposition and peer benchmark this quarter; adopt the national scorecard with no more than forty indicators tied to the four functions; legislate the annual state-of-resilience report; and rule, in standing orders, that no digital or resilience programme is approved without stating which indicator it moves — the sentence that converts measurement from commentary into steering.

The close: the first twelve months, through a Czech lens

A playbook that ends without a calendar is a wish. Here is the first-year sequence for a mid-sized EU state — written for the Czech Republic, portable to any of its peers — with the wiring named and the moves concrete. The premise throughout is the one the evidence base keeps repeating: Ukraine’s survival was prepared in the decade before February 2022 and enacted in days; Estonia’s data embassy went from concept to signed Luxembourg agreement because someone owned it; Finland’s model works because every vital function has a lead ministry and an exercise date. Ownership first, then motion.

The wiring. A National Resilience Council at the Office of the Government, chaired by the prime minister on the Finnish Security Committee pattern, with a small permanent secretariat — the centre-of-government unit that owns this playbook, the risk register, the exercise calendar, and the annual report to parliament. The Czech Statistical Office as data steward of the national map and the measurement layer, with new legal gateways to scanner, payments, and customs microdata. NÚKIB as the cyber floor’s owner — already the strongest institution on the board, per its own national strategy — extended with the agentic SOC mandate. The Ministry of the Interior and the integrated rescue system as crisis-operations owner; the Digital and Information Agency as owner of registers, identity, and continuity engineering; the Council for Research, Development and Innovation with the funding agencies as owner of the science roadmap; the Czech National Bank as nowcasting partner. One council, six named owners, one public scorecard.

The first moves.

1. Month 1 — stand up the Council and its secretariat, with a mandate letter assigning each of the six areas a named institutional owner and a twelve-month deliverable. No new ministry; a centre with convening power and a budget line.

2. Months 1–3 — publish the national risk register, on the UK model of acute risks with reasonable-worst-case scenarios, and audit the state against NATO’s seven baseline requirements and the four EW4All pillars. This is the gap map everything else prioritises against.

3. Months 2–4 — designate the continuity core: the ten registers and systems without which the Czech state cannot govern, on the Estonian ten-dataset pattern; begin engineering their replication and open negotiations for a data-embassy agreement with an allied host, Vienna Convention-style immunity included.

4. Months 2–5 — pre-draft the emergency instruments: the cloud-migration authorisation, emergency-procurement resolution, and data-sharing powers Ukraine had to improvise in the invasion’s first week, drafted now, debated calmly, and left ready for signature.

5. Months 3–6 — stand up the foresight-and-decision unit beside the strategy function at the Office of the Government: scanning agents running from day one, the OECD FIELD/SCOPES self-assessment completed, and two live policy questions — energy security and labour-market exposure to AI are the obvious Czech candidates — stress-tested against scenarios with ministers present.

6. Months 3–8 — build the first national dependency map: critical-functions taxonomy adopted, systemically important entities identified on the RAND method, and the first three EU-style in-depth reviews commissioned on the imports where Czech industry is most exposed; the Statistical Office and the central bank launch the scanner- and payments-data nowcast pilot in parallel.

7. Months 4–9 — close the cyber floor’s known gaps: NIS2/CER entity identification completed and published, CSF 2.0 profiles with named Govern-function owners mandated across ministries, the post-quantum inventory begun, and the GCI pillar profile answered gap by gap.

8. Months 6–10 — publish the research-infrastructure landscape analysis and the first national roadmap on ESFRI rules; designate two missions with Mazzucato-grade governance; secure EuroHPC AI-factory access and open a national research-compute scheme on the NAIRR pattern, with a white-space reserve written into the envelope.

9. Months 9–12 — run the national exercise: a compound scenario — cyber campaign plus infrastructure failure plus disinformation surge — with ministers in the room, operators at the table, the agentic common-operating-picture live, and the data-embassy restoration rehearsed. The exercise report goes to parliament unredacted wherever possible.

10. Month 12 — publish the first annual State of National Resilience report: the forty-indicator scorecard, the four-index benchmark, every red field paired with an owner and a funded fix — the document that turns the playbook from a programme into a habit.

Twelve months, no science fiction, nothing that a state of ten million cannot afford — most of it assembly of frameworks other institutions have already written, exercised by agent fleets that are already buildable, on data the state already holds. That is the closing argument, and it is an option-value argument. Building the four functions does not require knowing whether the next decade brings a war, a pandemic, a grid failure, or a quiet dependency trap. It requires deciding to be the kind of state that finds out early, decides quickly, knows what it has, and keeps discovering — the kind of state that can lose systems and still govern. The library’s verdict is that such states are made, not born: made in the unglamorous years before the crisis, by governments that treated resilience as a capability to be exercised rather than a report to be filed. The plan on the shelf saves no one. The rehearsed institution, wired to its data and its agents, with a human hand on every consequential decision — that is what sovereignty looks like when it is stressed. Build it now, while it is still cheap.

The argument, before the list

Sovereignty is not autarky. No serious government believes it can fabricate its own leading-edge chips, train its own frontier models, lay its own transoceanic cables and write its own operating systems — and the one audit that priced full self-sufficiency in even a single layer found it absurd: the European Court of Auditors concluded that in the microchip value chain “total autonomy is impossible”, before showing that even the EU’s far more modest 20% production target is very unlikely to be met (ECA, The EU’s Strategy for Microchips, 2025). The states that talk loudest about digital autarky tend to be the ones building censorship regimes, not capabilities. That is not the sovereignty this report is about.

Sovereignty, properly defined, is the ability to decide, act, and recover without asking permission — from another state, or from a foreign platform whose incentives are not yours. The EPRS puts it in one line: digital sovereignty is “Europe’s ability to act independently in the digital world” (EPRS, Digital Sovereignty for Europe, 2020). Fraunhofer ISI, in the paper that made “technology sovereignty” respectable as a policy concept, is careful to define it as a state’s capacity to provide or reliably access the technologies it deems critical — explicitly distinct from autarky and from protectionism (Fraunhofer ISI, Technology Sovereignty: From Demand to Concept, 2020). The academic mapping most cited on the subject makes the same move along three dimensions — the state, the economy, the individual — and finds that what governments actually mean by the word is control over their own digital environment, not isolation from everyone else’s (Internet Policy Review, Pohle and Thiel: Digital Sovereignty, 2020).

The reason this now belongs at the centre of statecraft rather than in a digital-ministry annexe is that the modern state has quietly become a tenant in its own house. Its registers sit in rented clouds, its communications cross cables it neither owns nor can repair, its chips arrive through a supply chain with single points of failure on the other side of the planet, its officials authenticate through platforms governed by foreign law, and its AI ambitions run on compute operated — even when physically located at home — by companies answerable to another jurisdiction. Each of these arrangements was individually rational. Renting is cheaper, faster and usually better-run than building. But the sum of rational rentals is a structural condition: the rented state can be switched off politely, contractually, completely — no invasion required, merely a sanctions decision, a licence withdrawal, a change in terms of service, or a distant boardroom concluding that your market is no longer worth the compliance risk.

This is not a hypothetical mechanism. The scholarship on weaponized interdependence has documented how the states and firms that sit at the hubs of financial, informational and technological networks convert that position into coercive leverage — surveillance of what flows through the hub, and the ability to cut adversaries off from it (CSET, From Cold War Sanctions to Weaponized Interdependence). And the accidental version is just as instructive as the deliberate one: the United Kingdom’s official risk register opens by citing the CrowdStrike IT outage as proof of how widely a single technology failure propagates through services essential to daily life (UK Cabinet Office, National Risk Register, 2025). A state can be partially switched off by a bad software update it never installed, in a product it never chose, running in systems it does not operate.

The actor in this report is therefore the state — not the firm. Firms optimise for efficiency and can exit a bad dependency by dying and being replaced; states cannot. Only the state carries the four functions that must survive stress: managing crises, making decisions under uncertainty, knowing what the nation has and depends on, and steering its scientific capacity toward its own problems. Those four functions stand on a stack of assets — physical substrate, compute, data, platforms, models, operations, people and rules — and for every layer of that stack the state faces the same three-way choice: own it, share it with allies, or rent it with a tested exit plan. All three are legitimate. What is not legitimate, after the evidence assembled here, is the fourth option most states have actually chosen: renting by default, without knowing it, with no fallback and no list of what has been rented.

The good news is that the map can be drawn. The best states already draw it — the UK enumerates 89 acute risks with reasonable worst-case scenarios; the US maps 55 national critical functions rather than a list of buildings; Estonia has decided precisely which ten datasets constitute its continuity as a state and backs them into an embassy abroad. This report walks the seven layers of the sovereign stack, names where the dependencies actually bite — with numbers — and ends with a ranked shortlist for a mid-sized EU state: what to own outright, what to federate at European level, and what to go on renting, eyes open, exit plan in hand.

The map in brief

• Sovereignty is a capability, not a slogan — the ability to decide, act, and recover without asking permission; it is explicitly not autarky, which even the richest blocs cannot afford in a single layer (ECA; Fraunhofer ISI).

• The state’s assets stack into seven layers — physical substrate; compute and cloud; data; platforms; intelligence and models; operational capabilities; people, rules and trust — and each layer demands an explicit own/share/rent decision.

• The dependencies are measured, not rumoured: 48% of 775 non-US data centre projects are operated by US companies when weighted by investment value (arXiv, How Sovereign Is Sovereign Compute, 2025); three hyperscalers hold a combined 70% of the EU cloud market (Clingendael, 2024); the EU’s own forecast puts its 2030 chip share at 11.7% against a 20% target (ECA, 2025); upwards of 95% of intercontinental internet traffic runs through roughly 475 submarine cables (Atlantic Council, 2021).

• The trap is mechanical, not moral. Lock-in is built from egress fees, closed APIs and path dependence (ACM Netherlands, 2022); interdependence becomes a weapon at the network hubs (CSET); the chokepoints sit in exactly three arenas of internet infrastructure — naming, routing, and cables (SWP, Cracks in the Internet’s Foundation, 2019).

• The best practice is to map functions, not assets — the shift CISA made with its National Critical Functions (2019), the UK made with its 89-risk register (2025), and the EU is forcing on member states through the CER and NIS2 regimes.

• Continuity can be engineered. Estonia’s data embassy — ten strategic registers replicated under Estonian legal control in Luxembourg — proves a state’s core can survive the loss of its territory’s infrastructure (e-Estonia, Data Embassy Factsheet); Ukraine’s wartime cloud migration proves it at war-scale (Atlantic Council, Building the Digital Front Line, 2025).

• For a mid-sized EU state the ranking is knowable: own the registers, identity, the data-exchange bus, continuity copies and crisis operations; federate compute, chips, space and standards at EU level; rent hyperscale cloud and frontier models — with contractual exits, tested annually.

• The first move is an inventory: no state can decide what to own until it has enumerated what it currently rents, from whom, under whose law, and what breaks first when it is withdrawn.

How this report is organised

The body follows the seven layers of the ENSI sovereign asset map, from the ground up — Layer 0, the physical substrate, through Layer 6, the human and governance assets — because each layer stands on the ones below it and inherits their fragilities. For each layer the treatment is the same: what the layer is, the documented evidence of dependency, the own/share/rent decision a state must make explicitly, and the signals that tell you the decision needs revisiting. After the seven layers comes a section on the mechanics of the dependency trap — how interdependence is weaponized and where the chokepoints sit — because the mechanism repeats across every layer and deserves to be understood once, properly. The close ranks the layers for a mid-sized EU state, with the Czech Republic as the home example.

1. Layer 0 — the physical substrate: energy, chips, cables, space, quantum

What it is. Everything digital is physical somewhere. Layer 0 is the matter under the stack: electricity for compute, semiconductors, the fibre and submarine cables data actually travels through, the space assets that provide positioning, timing, observation and fallback communications, and — arriving now — the quantum technologies that will reset both sensing and cryptography. A state that skips this layer builds its sovereignty on someone else’s ground.

Where the dependency bites. Semiconductors are the best-audited case. The EU’s share of global chip manufacturing stood at roughly 9% in 2020, and in 2021 — with its production sites at full capacity — the bloc still ran a trade deficit in microchips of almost €20 billion (ECA, The EU’s Strategy for Microchips, 2025). The Chips Act mobilised at least €43 billion in policy-driven investment, yet the auditors found the Commission directly responsible for only around 10% of the public funding and concluded the flagship 20%-by-2030 target is very unlikely to be met — the Commission’s own July 2024 forecast projects just 11.7% (ECA, 2025). The pandemic previewed the stakes: chip shortages collapsed German car production to 1975 levels (ECA, 2025). The chokepoints are structural, not cyclical — advanced fabrication concentrated in Taiwan and South Korea, lithography in one Dutch firm, design software in the US — a topology mapped supplier by supplier in CSET’s supply-chain brief (CSET, The Semiconductor Supply Chain). Cables tell the same story one layer down: upwards of 95% of intercontinental internet traffic crosses roughly 475 submarine cables, whose ownership is shifting toward large internet platforms and, on some routes, Chinese state-linked builders, while remote management systems create new single points of operational failure (Atlantic Council, Cyber Defense Across the Ocean Floor, 2021). The European Parliament’s own assessment reaches the equivalent conclusion for EU connectivity (EP Policy Department, Security Threats to Undersea Communications Cables, 2022), and Taiwan is the live experiment: an island economy of world-systemic importance connected through 15 undersea cables, with documented sabotage pressure on them (Stanford FSI, Taiwan Undersea Cable Network Resilience, 2024; DSET, Vulnerabilities at Depth, 2025). In space, Europe has drawn the lesson from Ukraine’s dependence on a single private constellation and is building IRIS² as a sovereign secure-connectivity layer (ESPI, IRIS² Growing Up, 2025) — while the space-cyber intersection becomes its own attack surface (ESPI, Space Cyber and Defence, 2023). Quantum sits at the edge of the layer: RAND’s sober reading of commercial and military timelines argues against panic and for sequencing (RAND, Applications and Timelines for Quantum Technology) — but post-quantum cryptographic migration is a defensive asset a state must start owning now, because harvested traffic decrypts retroactively.

The own/share/rent decision. No mid-sized state should own a leading-edge fab — the ECA audit is effectively a proof that even the EU cannot buy its way to one-fifth of the market. The rational portfolio is: own the dependency map (which chips, from whom, substitutable where), own stockpiles for the chips its critical infrastructure actually consumes, and own niche positions where it has genuine comparative advantage — design, packaging, materials, instruments. Share the rest at EU level: EuroHPC for the big iron, the Chips Act for what it can realistically deliver, IRIS² and Galileo for space. Connectivity is the exception that demands ownership: internet exchange points on national territory, redundant cable routes, vetted 5G/6G vendors, and — the most neglected asset in Europe — guaranteed access to cable repair capacity, which Taiwan’s analysts identify as the binding constraint in every sabotage scenario (Stanford FSI, 2024). Energy for compute must be owned as a matter of grid planning: it is the input every layer above silently assumes.

What to watch. The EU’s actual (not announced) fab investment against the ECA’s 11.7% trajectory; cable-laying and cable-repair fleet ownership by flag; the ratio of national IXP-routed traffic to traffic that hairpins through foreign hubs; IRIS² deployment against schedule; national PQC migration deadlines — and whether your own government has one.

2. Layer 1 — compute and cloud: the most rented layer in the stack

What it is. The machines the state’s digital life actually runs on: the cloud estates hosting government workloads, the AI compute national ambitions depend on, and the continuity copies that let a state survive the loss of its own data centres. This is the layer where the gap between the rhetoric of sovereignty and the accounting of it is widest — because it is the layer states have rented most completely, and most invisibly.

Where the dependency bites. Start with the number that should reframe every “sovereign cloud” announcement in Europe: an audit of 775 data centre projects outside the United States found that 48% are operated by US companies when weighted by investment value — and that is on top of the estimated 54% of worldwide compute capacity that sits inside US borders to begin with (arXiv, How Sovereign Is Sovereign Compute, 2025). The paper’s conclusion is the uncomfortable one: building data centres on national soil does not confer sovereignty if the operator answers to another legal system, because jurisdiction follows the operator’s nationality as well as the facility’s territory. A domestically sited hyperscaler region is, legally, a foreign object on home ground — this is the CLOUD Act problem in physical form. The market structure compounds the legal one. In the EU, three American providers hold a combined 70% of the cloud market (Clingendael, Too Late to Act? Europe’s Quest for Cloud Sovereignty, 2024); the Dutch competition authority found Microsoft Azure and Amazon Web Services each holding 35–40% of the IaaS and PaaS layers in the Netherlands and in Europe, with Google a strong third (ACM Netherlands, Market Study Cloud Services, 2022). And the lock-in is engineered, not incidental: the ACM documents how free ingress and expensive egress, closed APIs, complex tariff structures and deep service interconnection create path dependence from the first moment of choice — the initial procurement decision is effectively the permanent one unless exit is designed in from the start. Clingendael’s framing is the right strategic register: Europe is living its “5G moment” on cloud — the same slow realisation it went through with Huawei, one layer up, with the dependency this time running through allied rather than adversarial firms, which changes the threat model but not the structure. Demand for something better is real — Capgemini’s survey of a thousand organisations found sovereignty concerns now shaping cloud procurement across Europe (Capgemini, The Journey to Cloud Sovereignty, 2022) — but supply has lagged: Gaia-X chose to build a federation and trust framework rather than a European hyperscaler (Gaia-X, Architecture Document, 2022), and a decade of EU digital-sovereignty initiatives has produced, in the words of the Commission’s own commissioned stock-take, a “convoluted journey” (IDC, Digital Sovereignty in the EU, 2025).

AI compute sharpens all of it. Compute is the most governable input to AI — physical, detectable, concentrated in its supply chain — which is precisely why it is becoming an instrument of policy between states (arXiv, Computing Power and the Governance of AI, 2024). A state with no public AI compute has outsourced not just a workload but the option to have an AI strategy at all. The pattern responses exist and are documented: the OECD’s blueprint tells governments to plan national compute the way they plan energy — measuring capacity, effectiveness and resilience, not just buying GPUs (OECD, Blueprint for Building National Compute Capacity for AI, 2023); the UK ran a national needs assessment and stood up the AI Research Resource on the back of it (UK Government, Independent Review of the Future of Compute, 2023); the US designed the NAIRR to democratise access to research compute as public infrastructure (NSF, NAIRR Task Force Final Report, 2023); and EuroHPC is the working proof that mid-sized states can co-own big iron none could afford alone, now extending into AI factories and quantum machines (EuroHPC JU, Multi-Annual Strategic Programme 2021–2027).

The own/share/rent decision. Own three things outright: a protected enclave for the workloads whose exposure to foreign jurisdiction is intolerable — registers, security, justice, crisis systems; the continuity copies of the state’s critical systems, held under the state’s own legal control (the data embassy pattern, of which more in Layer 2); and the contractual and technical machinery of exit — portability tested annually, egress priced into every contract, no critical workload without a rehearsed landing zone elsewhere. Share AI compute at European level through EuroHPC and its AI factories, buying national slices of shared capacity. Rent the hyperscale bulk — it is genuinely better run than anything a ministry will build — but rent it knowingly: classified workload tiers, jurisdiction mapped per workload, and the ACM’s lock-in mechanics treated as a checklist of what to contract away. The unglamorous prerequisite for all of it is data classification: Clingendael identifies proper classification of government data — knowing which workloads are sovereignty-critical and which are commodity — as one of the two capabilities European governments most conspicuously lack, alongside the talent to manage hybrid estates (Clingendael, 2024). A state that cannot classify its workloads cannot tier its cloud, and defaults to treating everything as commodity — which is how registers end up next to newsletters. The test of cloud sovereignty is not where the servers sit; it is whether the state can move, and has proven it can.

What to watch. The operator-nationality share of new national data centre capacity (the How-Sovereign metric, tracked annually); egress fees and portability obligations as the Data Act bites; the ratio of state workloads with a tested exit plan to those without — a number almost no government currently knows about itself; EuroHPC AI-factory capacity actually allocated to your national researchers and firms.

3. Layer 2 — data: the registers are the crown jewels

What it is. The state’s knowledge of its own nation: the base registers — population, property, business, address, vehicle — that every other function reads from; the statistical system; the geospatial and infrastructure maps; health, education and welfare records; and the economic micro-data that reveals dependencies before they bite. If Layer 1 is where the state runs, Layer 2 is what the state is, informationally. A state that cannot enumerate its people, firms and land cannot tax, mobilise, respond, or rebuild — whatever else it still controls.

Where the dependency bites — and where the proof case lives. The canonical framework here is the World Bank’s: data as a national asset governed under a “social contract for data”, with the infrastructure, institutions and safeguards that make reuse possible without destroying trust (World Bank, World Development Report 2021: Data for Better Lives, 2021). The OECD maps the same terrain across member states — openness against control, and the governance machinery that lets a state have both (OECD, Going Digital to Advance Data Governance, 2022). National strategies show what taking the layer seriously looks like: the UK’s five-mission strategy treats government’s own data use as sovereign capability, not administrative plumbing (UK Government, National Data Strategy, 2020); Germany’s Datenstrategie runs to some 240 measures, from data trusteeship to state modernisation (German Federal Government, Data Strategy, 2021); the EU is building sectoral data spaces — health, energy, mobility, public administration — as shared European infrastructure with common governance (Council of the EU, Common European Data Spaces, 2022). And because data crosses borders even when law does not, UNCTAD’s mapping of the world’s divergent data-flow regimes is the reminder that this layer is foreign policy as much as domestic administration (UNCTAD, Digital Economy Report 2021).

But the document that should be pinned above every CIO’s desk is two pages long. Estonia — the state that has thought hardest about digital continuity because its history obliges it to — has enumerated precisely ten strategic datasets whose survival constitutes the survival of the state: the population register, the business register, the land and cadastral registers, the identity documents register, the taxable person’s register, the treasury information system, the court e-file, the State Gazette, and the national pension insurance registry. These are continuously replicated to a data embassy in Luxembourg — a Tier 4 data centre that is legally Estonian territory in the relevant senses, fully under Estonian control, with the immunities of a physical embassy, under an agreement signed in 2017 (e-Estonia, Data Embassy Factsheet; European Commission, Estonia Data Embassy Initiative Case EE05). Read that list carefully: it is the revealed-preference answer to this report’s central question. Asked “what must a state actually control?”, the state that has thought longest answered with ten registers — not a fab, not a search engine, not a social network. The registers are the crown jewels; everything else in the stack exists to serve, protect, move and reason over them.

One class of data deserves separate emphasis because almost no state collects it deliberately: economic micro-data — firm-level supply chains, customs flows, payments telemetry. This is the data that would have shown the chip chokepoints before they collapsed German car production to 1975 levels (ECA, 2025), and it is the raw material for the strategic-dependencies mapping the EU now attempts at Union level and RAND has methodised at entity level (RAND, Identifying and Prioritizing Systemically Important Entities, 2023). A state that cannot see which of its firms depend on which foreign inputs is running its economy on the same blind trust it has been running its cloud — and this is the dataset Layer 4’s simulations and early-warning agents will starve without.

The own/share/rent decision. This is the one layer where the answer is nearly absolute: own it. Base registers, the statistical system and the authoritative geospatial map must sit under national legal control without exception — including their backups, which is the data embassy’s real lesson: ownership of the primary copy means little if continuity depends on infrastructure that falls with the territory or sits under foreign law. Share at European level where sharing multiplies value without transferring control — the data spaces, cross-border register interoperability, statistical standards. What may be rented is processing capacity underneath the data, under the Layer 1 rules — never the stewardship, the schema, or the legal custody of the registers themselves. Health and welfare records add the trust constraint: the WDR’s social contract is operational guidance here, because a population that stops trusting the state with its data will stop feeding the registers, and the asset degrades from within.

What to watch. Whether your state can produce its own ten-dataset list — and how long the argument over its contents takes, which measures how little the question has been asked; the existence, location and legal instrument of continuity copies for each register on the list; time-to-restore, exercised rather than asserted; administrative-data pipelines replacing surveys in the statistical system — the difference between quarterly hindsight and operational awareness; and the share of critical registers whose processing sits with operators under foreign jurisdiction, which quietly re-imports the Layer 1 problem into the layer the state believed it owned.

4. Layer 3 — platforms: identity, payments, exchange, and the state’s front door

What it is. The digital public infrastructure through which citizens and firms actually meet the state: a guaranteed way for every person and business to authenticate and sign; payment rails; a secure data-exchange bus connecting registers and agencies; the trusted channels through which the state speaks and delivers; and the open-source capacity to read, fork and maintain the software all of it runs on. Layer 2 is what the state knows; Layer 3 is how that knowledge becomes services, and how the state remains present in its citizens’ lives when everything else is mediated by platforms it does not control.

Where the dependency bites. The platform layer is where the sharpest sovereignty question hides in the friendliest packaging: if identity, payments and communication between citizen and state run through private foreign platforms, the state has outsourced the relationship itself — and with it the ability to reach every citizen in a crisis, to guarantee a firm can transact, to know its channels will exist next year on the same terms. The global policy consensus has now converged on the answer: identity, payments and data exchange are the three building blocks a state must guarantee as public infrastructure — the G20-endorsed DPI framing (UNDP, Accelerating the SDGs Through Digital Public Infrastructure, 2023), with Carnegie’s thesis making the design point that states must own the rails and the rules while markets build on top (Carnegie Endowment, The Future of Digital Public Infrastructure). The proof cases are no longer theoretical. India built identity and payments as public rails — Aadhaar and UPI — and the BIS’s analysis of the India Stack is the standard reference for what sovereign digital financial infrastructure at population scale looks like (BIS, Design of Digital Financial Infrastructure: Lessons from India, 2019). Estonia built the exchange layer: X-Road, the secure data-exchange bus that turns a pile of siloed registers into a networked state, now exported worldwide (e-Governance Academy, X-Road Secure Data Exchange Concept, 2022; e-Governance Academy, e-Estonia: e-Governance in Practice). Singapore built the delivery machine — strategic national projects run by a dedicated engineering agency rather than procured wholesale (Smart Nation Singapore, The Way Forward, 2018). Europe is now building identity at continental scale through the EUDI wallet’s common architecture (European Commission, EUDI Wallet Architecture and Reference Framework), on assurance foundations set by NIST’s identity guidelines — the IAL/AAL/FAL framework underpinning government digital identity worldwide (NIST, Digital Identity Guidelines SP 800-63-3) — with the World Bank’s ID4D guide as the institutional playbook (World Bank, ID4D Practitioner’s Guide, 2019) and GovStack packaging the whole pattern as reusable building blocks for any state (ITU, GovStack and Digital Public Infrastructure).

The own/share/rent decision. Own the three rails — identity, the exchange bus, and at least a public option in payments — plus the state’s crisis-grade channel to every citizen. Ownership here means the state guarantees the rail, sets its rules and holds its keys; building and operating can be contracted, but the rail must not be a private product the state merely uses. Share the standards and wallets at EU level — eIDAS and the EUDI wallet are exactly the right altitude, one specification, twenty-seven sovereign implementations. Rent the commodity underneath — app development, hosting under Layer 1 rules — and cultivate the open-source escape hatch deliberately: the capacity to read, fork and maintain critical software is what converts an unexitable vendor lock into a negotiation, and procurement is the lever that builds it. What to watch: EUDI wallet adoption and the share of high-value services accepting it; whether a public instant-payments option exists and merchants actually route through it; the number of registers connected to the exchange bus versus point-to-point integrations; the share of critical government software the state could fork and maintain if it had to.

5. Layer 4 — intelligence and models: the state’s ability to ask “what if?”

What it is. The reasoning layer: access to frontier AI models and sovereign fine-tuned models for the state’s languages, laws and classified domains; simulation and digital twins of the economy, infrastructure and territory; nowcasting and early-warning feeds that shorten the state’s observation loop from quarters to days; and an institutionalised foresight capability wired into budgeting rather than shelved in reports. Layers 0–3 determine whether the state can act; Layer 4 determines whether it acts intelligently — whether it can ask “what happens if?” before reality answers on its own schedule.

Where the dependency bites. This layer inherits every dependency below it and adds one of its own: model dependency. A state whose analytical stack runs on rented frontier models has rented a component of its own judgement — the models’ availability, pricing, refusals and failure modes are all set elsewhere, and the compute-governance literature makes clear that access to AI capability is becoming an explicit instrument of interstate policy, granted and withdrawn like any other strategic export (arXiv, Computing Power and the Governance of AI, 2024). The concentration of what AI Now calls infrastructural power — compute, models, and the platforms in one set of corporate hands — means the state is not one customer among many but a dependent among giants (AI Now Institute, 2023 Landscape: Confronting Tech Power, 2023). The answer is not a national frontier lab — for a mid-sized state that is the fab fallacy one layer up. It is a portfolio: rented frontier access for the general case, sovereign fine-tuned models for the domains where language, law and classification make foreign models unusable, and — the part almost every state still lacks — the simulation and early-warning assets that no vendor sells off the shelf because they must be built on the state’s own Layer 2: a digital twin of the production network fed by firm-level micro-data, grid and epidemic nowcasts fused from the state’s own sensors, scenario machinery connected to the risk register rather than to a shelf.

This is where the agentic engine becomes the design pattern. The functions of Layer 4 are precisely the ones AI agents can now run continuously rather than annually: scanning agents sweeping signals against the risk register; simulation agents keeping the national digital twin warm and running counterfactuals overnight; early-warning agents watching the dependency map — operator nationality, cable repairs, chip inventories, register anomalies — and escalating on threshold; red-team agents attacking the state’s own assumptions; briefing agents compressing all of it into the morning’s decision documents. Humans own judgement and accountability; the agents own the cadence. A foresight function staffed only by humans reports quarterly; a foresight function run as an agentic engine reports continuously — and the difference is measured in the length of the state’s observation loop, which is the quantity this whole layer exists to shorten.

The own/share/rent decision. Own the twins, the nowcasts, the early-warning wiring and the foresight institution — they are made of your own data and are worthless rented. Own fine-tuned models for law, administration and the classified domain. Share evaluation infrastructure, safety testing and model-access agreements at EU level, where collective weight buys terms no mid-sized state gets alone. Rent frontier capability — with more than one supplier, and with the fine-tuned sovereign fallback tested against the day rented access is degraded, priced up, or withdrawn. What to watch: the state’s observation loop, function by function — days or quarters; whether a national digital twin of the production network exists and is fed continuously; the share of government AI workloads that would survive the loss of a single model vendor.

6. Layer 5 — operations: cyber defence, crisis machinery, and the standing map

What it is. The layer where sovereignty is exercised rather than possessed: national cyber defence covering the critical sectors; crisis-management systems — a common operating picture, resilient communications, population warning, rehearsed continuity plans; the decision infrastructure of the centre of government; science and innovation planning; and the continuous mapping of the nation’s assets and dependencies. Everything above Layer 5 is inventory; Layer 5 is practice.

Where the evidence points. On cyber defence the doctrine has converged across the Atlantic: the US strategy rebalances responsibility for security away from end users and toward the actors most capable of bearing it, using regulation and market incentives (White House, National Cybersecurity Strategy, 2023), executed by CISA’s plan to harden critical infrastructure sector by sector (CISA, Cybersecurity Strategic Plan FY2024–2026, 2023); the UK runs a whole-of-government hardening programme to 2030 (UK Cabinet Office, Government Cyber Security Strategy 2022–2030); the EU drags every essential and important entity up to a common floor through NIS2 (EPRS, The NIS2 Directive, 2021), with ENISA’s threat landscape naming the actors and techniques the floor is being raised against (ENISA, Threat Landscape 2024) and its Union-wide assessment measuring how unevenly member states are climbing (ENISA, State of Cybersecurity in the Union, 2024). NIST’s CSF 2.0 added a Govern function to the world’s reference framework — the standards body’s way of saying cyber risk is now a board and cabinet matter, not an IT matter (NIST, Cybersecurity Framework 2.0, 2024). And a mid-sized state need not imagine what national capability looks like at its scale: the Czech Republic’s NUKIB strategy is a working model of small-state cyber sovereignty — national CERT coverage, regulated critical sectors, exercised response (NUKIB, National Cyber Security Strategy 2021–2025).

On crisis machinery, three exemplars define the standard. Estonia 2007 is the founding case — the first state-scale cyber campaign, and the origin of the lesson that resilience is an information-warfare posture, not an IT department (CCDCOE, Estonia 2007 Cyber Attacks Analysis, 2008). Ukraine 2022 is the modern proof: pre-scripted legal changes and a wartime migration of state systems to cloud infrastructure outside the country kept the government operating under physical attack — digital continuity as a survival capability, improvised well only because it had been prepared (Atlantic Council, Building the Digital Front Line, 2025). Finland is the peacetime model: comprehensive security as a whole-of-society doctrine, with vital functions assigned, exercised and drilled across government, business and citizens (Finland Security Committee, Security Strategy for Society, 2017).

On the standing map — the function ENSI regards as the quiet core of the layer — the methodological shift is documented in Layer 6’s own sources: from listing assets to mapping functions (CISA’s 55 National Critical Functions, 2019), from protection to resilience (OECD, Good Governance for Critical Infrastructure Resilience, 2019), from static lists to ranked systemic importance (RAND, Identifying and Prioritizing Systemically Important Entities, 2023), all published transparently enough to steer the private owners of the infrastructure (UK Cabinet Office, National Risk Register, 2025). The EU’s CER regime exists precisely because most member states had not done this work: under the old directive just 94 European critical infrastructures were designated across the whole Union — two-thirds of them in three member states, and sixteen member states had designated none at all (EPRS, Improving the Resilience of Critical Entities, 2021). A map that two-thirds empty is not a map; it is an assumption.

The own/share/rent decision. Own all of it — this layer is the definition of the state’s job, and renting incident response, crisis coordination or the national risk map means renting government itself. What can be shared is intelligence and exercises: threat intelligence through ENISA and allied CERTs, crisis exercises with neighbours, mutual-aid response capacity. What can be bought is tooling and surge expertise — under contracts that survive the crisis they will be invoked in, which is exactly the clause worth testing before it is needed. What to watch: exercise cadence — when the state last rehearsed a register restore, a cloud exit, a cable cut, a cell-broadcast to the whole population; NIS2/CER designation coverage against the 94-entity baseline; mean time from national incident to common operating picture, which is the single best proxy for whether Layer 5 exists at all.

7. Layer 6 — people, rules, and trust: the assets that cannot be procured

What it is. The top of the stack is not technology. It is a digitally capable civil service with retained in-house engineering, so the state is a competent buyer and builder rather than a hostage of its integrators; surge-capable reserves of technical talent that can be mobilised in crisis; seats at the tables where the defaults of the digital world are written; legal capacity to act fast under oversight; and the trust of a population that will believe official channels when it matters. Every other layer can, in principle, be bought or federated. This one compounds or decays, and does either slowly.

Where the evidence points. The delivery-capable states are distinguishable by one organisational fact: they kept engineering inside government. Estonia’s digital state was built by a state that could build (e-Governance Academy, e-Estonia: e-Governance in Practice); Singapore’s Smart Nation runs on a government engineering agency executing strategic national projects directly (Smart Nation Singapore, The Way Forward, 2018); and Finland’s comprehensive-security model treats trained people across society — not agencies — as the resilience asset itself (Finland Security Committee, Security Strategy for Society, 2017). Rules are made in rooms, and absence from the rooms is a dependency like any other: the governance of naming, routing and standards is contested terrain where authoritarian states push for a state-controlled internet through the ITU while the multi-stakeholder bodies drift toward politicisation — a state with no standards capacity simply inherits whichever defaults win (SWP, Cracks in the Internet’s Foundation, 2019). And trust is the layer’s load-bearing wall. Global internet freedom has been declining for years as states normalise digital controls (Freedom House, Freedom on the Net 2024) — which sets the trap precisely: a state that defends its information space by closing it destroys the trust the defence was meant to protect, while a state that ignores the information layer entirely finds, as Estonia did in 2007, that trust is exactly what a hostile campaign targets (CCDCOE, Estonia 2007 Cyber Attacks Analysis, 2008). The narrow path — openness plus institutional credibility plus channels the population actually believes in a crisis — must be built in peacetime, because it cannot be improvised under attack.

The own/share/rent decision barely applies — people, standards seats and trust cannot be rented, which is the point. The actionable form: pay the salary premium for a core in-house engineering cadre and treat it as critical infrastructure; build the reserve model before the crisis; fund standards participation as a strategic expense, pooling with EU partners where one delegation can carry twenty-seven states’ interests; and treat every crisis communication as a deposit in, or withdrawal from, the only account that matters in the worst week. What to watch: the ratio of state engineers to contractor engineers on critical systems; national names on standards-body rosters; measured trust in official channels — tracked, like any other strategic stock, before it is needed.

8. The mechanics of the trap — how rented layers become leverage

The seven layers share one failure mode, and it deserves to be understood once, precisely, because it repeats at every altitude of the stack.

Interdependence is weaponised at the hubs. The networks the digital economy runs on — financial messaging, cloud platforms, chip supply chains, cables — are not flat; they have hubs, and whoever controls a hub gains two powers over everyone routed through it: the ability to watch what flows, and the ability to cut adversaries off. That is the weaponized-interdependence thesis, and the sanctions record shows the machinery being used deliberately and repeatedly (CSET, From Cold War Sanctions to Weaponized Interdependence). The strategic consequence inverts the free-trade intuition: integration into someone else’s network is not neutral efficiency — it is exposure whose price is set later, by the hub.

The chokepoints are few and known. For the internet’s own infrastructure they sit in exactly three arenas — the naming system, the routing system, and the submarine cables — each already an active site of geopolitical conflict, with authoritarian states working to relocate authority over them into bodies they can dominate (SWP, Cracks in the Internet’s Foundation, 2019). The cable arena shows how concentrated a “distributed” system really is: upwards of 95% of intercontinental traffic through some 475 cables, with ownership consolidating into the same few platforms that dominate the layers above, and remote management systems adding a new class of single points of failure (Atlantic Council, Cyber Defense Across the Ocean Floor, 2021). Taiwan is the stress case the rest of the world should read as its own future perfect: fifteen cables, documented grey-zone pressure against them, and repair capacity — not cable count — as the binding constraint (Stanford FSI, 2024; DSET, Vulnerabilities at Depth, 2025).

The micro-mechanics are contractual, not dramatic. Most leverage is never exercised as a cut-off; it is exercised as terms. The ACM’s cloud study is the anatomy: free ingress and priced egress, closed APIs, tariff structures too complex to compare, services engineered to interlock — so that the cost of leaving compounds silently while the cost of staying is always, this quarter, smaller (ACM Netherlands, Market Study Cloud Services, 2022). Multiply that by the market structure — two firms at 35–40% each of the layers everything else runs on, 70% across three (ACM, 2022; Clingendael, 2024) — and by the legal reach that follows the operator across borders (arXiv, How Sovereign Is Sovereign Compute, 2025), and the trap needs no villain. Every actor optimises locally; the dependency assembles itself.

And the leverage runs through allies. The uncomfortable core of the European evidence is that the dependency is overwhelmingly on friendly firms under friendly law — nine of the ten largest GDPR fines have landed on American Big Tech (Clingendael, 2024). Alliance lowers the probability that leverage is used; it does not remove the leverage, and the probability is not under your control — which is why the asset map, not the alliance, is the correct planning basis. States that map their dependencies can price them; states that do not will discover them, as the pandemic’s chip shock and the CrowdStrike outage demonstrated, at the moment of maximum cost (ECA, 2025; UK Cabinet Office, 2025).

What to do first — a ranking for a mid-sized European state

For a state of the Czech Republic’s scale — ten million people, an open economy inside the EU, a capable cyber authority already on the field (NUKIB, 2021–2025) — the seven layers resolve into three lists and a sequence.

Own outright — the sovereign minimum. The base registers and the statistical system, with continuity copies under national legal control in the data-embassy pattern — a state that cannot lose its registers cannot lose its statehood to a server failure. Digital identity, the data-exchange bus, and a crisis-grade channel to every citizen. The national CERT/SOC, the crisis common operating picture, and the standing asset-and-dependency map, run as a continuous function on the NCF/NRR pattern — 55 functions and 89 risks are not American and British numbers; they are proof the enumeration can be done and published. A core in-house engineering cadre. Internet exchange points on national territory and a tested national PQC migration plan. None of this requires frontier technology; all of it requires the decision that it will not be rented.

Federate at EU level — the shared tier. AI compute through EuroHPC and its AI factories; chips through the Chips Act read at the ECA’s realistic altitude — niches, stockpiles and mapped dependencies, not fabs; space through Galileo, Copernicus and IRIS²; identity standards through eIDAS/EUDI; data spaces; standards-body presence pooled with like-minded delegations. The federated tier is not second-best sovereignty — it is the only arithmetic under which mid-sized states get these layers at all.

Rent with a tested exit — the eyes-open tier. Hyperscale cloud for everything outside the protected enclave — under contracts that price egress, mandate portability, and are exercised annually with an actual workload actually moved. Frontier AI via more than one vendor, with sovereign fine-tuned fallbacks for law, administration and the classified domain. Commodity software everywhere, with the open-source fork capacity that turns lock-in into negotiation.

The sequence is a year’s work. First, the inventory: one register of digital dependencies — every critical workload, its operator, its jurisdiction, its exit plan or absence of one — built the way the UK builds its risk register and maintained by a standing cell, not a one-off consultancy. Instrument it with the agentic engine: scanning agents watching the dependency map and the world’s disclosure feeds, early-warning agents escalating on threshold, red-team agents attacking the state’s own continuity assumptions, briefing agents putting one page in front of the cabinet each week — humans owning every judgement, agents owning the cadence no human staffing level can sustain. Second, the ten-dataset decision: name the registers whose survival constitutes the state’s survival, and stand up the continuity copies — Estonia has published the template; copying it is a legal agreement and a data centre away. Third, the first full exercise: restore a register, exit a cloud, lose a cable, brief the public — and let the failures write year two’s budget.

The option-value argument closes where the stack began. A state cannot know which layer will be stressed first — a cable, a vendor, a fab, an update, a campaign against trust itself. What it can know is its own map: what it owns, what it shares, what it rents, and what happens next when any rented layer is withdrawn. Sovereignty in the digital age is exactly that knowledge, held current, exercised annually, and priced into every procurement — a capability, not a report. A state sovereign in this sense can lose systems and still govern. A state that has rented all seven layers without noticing can be switched off — politely, contractually, and completely — and the whole purpose of the map is to make sure that sentence is never discovered to be a description.

That world has ended. Not slowly, not partially — ended. Knowledge is now abundant, instant and nearly free; any child with a phone holds more of it than any library of the last century, and a machine beside them can reason over it, in any language, for the price of electricity. The transmission engine has been made obsolete by a better one. And so the institution built around the old scarcity is now optimising, with tremendous discipline and expense, for the one thing the world no longer needs: humans who can store and reproduce information.

Say that plainly, in a faculty lounge or a ministry hearing, and you will hear the verdict come back fast. Unproven. Radical. Fine for a TED talk, not for a syllabus. You cannot build a curriculum on “sovereignty” — where is the standard, where is the test, where is the evidence it will get a child into a good university or a good job? Attention and judgement and taste are not subjects; they are things a well-run school produces as a side effect, if it produces them at all, and no accreditation board will sign off on teaching them instead of the fundamentals. Parents did not send their children here to be made into philosophers. Teach the fundamentals. Teach what can be measured. Leave the rest to chance, as it has always been left to chance.

Guilty as charged, on every count but the last. We are proposing to reorganise the whole of school around capacities the standard curriculum treats as extracurricular at best. We are saying that the measured, tested, timetabled things — the spelling lists, the date-recall, the procedural arithmetic a calculator did decades ago and a language model now does in every domain at once — are no longer the point, if they ever fully were. We plead guilty to radicalism, and we file, as our defence, the single fact the accusation quietly assumes away: the old curriculum was never neutral. It was an engineering choice, made for an economy that has since been replaced, and defending it now as “the fundamentals” is not caution. It is nostalgia wearing the mask of rigour. The only real question is not whether to change the curriculum. It is whether we change it before or after this generation pays for our hesitation.

We hold this to be self-evident: that when a capacity becomes abundant and nearly free, the wise response is not to keep drilling it as though it were still scarce, but to move the object of education to the capacity that is not free — the one abundance cannot manufacture, cannot automate, and cannot devalue, because it lives nowhere but inside a human being. That capacity is not what a person knows. Machines now know more, and know it faster, and forget nothing. It is what a person is — the ability to concentrate, to discern truth from plausible noise, to originate rather than only reproduce, to judge well, to lead, to take responsibility for outcomes, to want the right things and to know why. None of this is on the timetable. All of it is about to be the whole of the game.

This is a manifesto, so it will not hedge past that point. The school we have is not merely outdated; it is actively miseducating a generation for a world that will not exist, and it is doing so at exactly the moment when getting it right has never mattered more. The stakes are not test scores or economic competitiveness, though those are real and though we will not pretend otherwise. The stakes are whether the next generation walks into the age of abundant intelligence as its owners or as its subjects — as the humans who direct the machines, or the humans the machines direct. We choose our words for this on purpose. The language of these sixteen theses is the language of ownership, sovereignty and power, not the soft managed vocabulary of the education committee, because the old vocabulary produced the old school, and a different institution needs a different theory of the person it exists to build: not a vessel to be filled, not a resource developed for the labour market, but a sovereign — one who owns their attention, their judgement, their capital, their meaning and their choices, and who therefore cannot be owned by anyone or anything else. We value the owner over the employee, the built faculty over the filled memory, the sovereign over the vessel — not as a rejection of knowledge, discipline or rigour, but as the honest redirection of all three toward what is actually scarce now.

The machines are going to be extraordinary. That is now certain, and no thesis below argues otherwise. The only open question — the question these sixteen theses exist to answer, and the question every parent, teacher and policymaker reading this is already implicated in, whether they have chosen a side or not — is whether the humans standing beside the machines will be extraordinary too, or whether we will have spent their one irreplaceable childhood preparing them, with great care and at great expense, for a race they were always going to lose. They draw on a purpose-built ENSI library of 179 primary studies across 30 research angles — the neuroscience of attention, the science of learning, the economics of automation, the psychology of motivation and meaning — but a manifesto is not a literature review, and what follows is not a summary of that evidence. It is a set of convictions, each grounded in it, about what we owe the children about to be sent into the most transformed world in human history. They are meant to be argued with. They are meant to be acted on. And they begin, as any honest declaration must, with a fact about the world before they arrive at a demand about what to do with it.

The sixteen theses, in short

1. Intelligence is now abundant; the human is the scarce ingredient — and value has always flowed to the scarce thing, without exception, without appeal.

2. The school we have optimises for the losing side of the race — it drills, with real discipline, the exact cognition we just finished automating.

3. Education’s purpose is to build a sovereign, not to fill a vessel — a person who owns their mind, not one stocked with facts they cannot yet use.

4. What is automated is devalued; what is human appreciates — this is the law under the whole transformation, and it tells you exactly what to teach.

5. AI is the greatest tutor and the greatest cognitive off-switch at once — the technology never decides which one you get; the design always does.

6. Struggle is not the obstacle to learning; struggle is the learning — remove it in the name of kindness and you remove the growth along with it.

7. Attention is the master resource — whoever owns it owns everything built downstream of it; whoever loses it owns nothing at all, however much they know.

8. Knowledge is not obsolete; it is the precondition for everything AI makes valuable — “just look it up” is not a shortcut, it is a cognitive-science error.

9. The scarce skills are judgement, taste, agency and meaning — none of them appear on any test we currently give, and that is the emergency, not a detail.

10. Teach with AI and defend the mind from it — both, at once — a school that runs only one mandate produces a cripple, however well-intentioned.

11. The goal is the owner, not the employee — the protagonist who acts on their own initiative, not the extra who waits, however excellently, to be told.

12. Money, capital and ownership must be taught to everyone — or the ownership divide widens quietly, unopposed, generation after generation.

13. Meaning is the emergency of an age of abundance — a school that treats it as enrichment rather than infrastructure is manufacturing despair at scale.

14. The human must hold the reins — accountability is the one property a machine can never inherit, no matter how capable it becomes.

15. Every one of these faculties is trainable — the barrier to building them was never feasibility. It has only ever been habit.

16. Build sovereigns — or hand the next generation a world they can only serve — there is no third, neutral option, whatever the timetable pretends.

The sixteen theses

1. Intelligence Is Now Abundant; the Human Is the Scarce Ingredient

Value has never once, in the history of markets, flowed to the abundant thing. It has no reason to start now.

Metaphor: When water is scarce it is sold by the drop; when it floods the streets, the person who can channel it is worth more than all the water combined.

We hold:

• For all of history, intelligence was the bottleneck on everything worth doing.

• It was rare, expensive, and lodged in a small number of scarce human experts.

• Machine intelligence has broken that scarcity, and nothing suggests it will return.

• We reject the premise that abundant intelligence keeps its old market value out of habit.

• We hold instead that value migrates, on schedule and without sentiment, to whatever remains scarce.

• The human who directs intelligence, not the human who merely possesses some of it, is that scarce thing now.

Why it holds:

• The OpenAI exposure study finds roughly 80% of workers have tasks exposed to large language models — cognitive labour commoditised at scale, not in theory but already, this year.

• Autor’s and Brynjolfsson’s complementarity work shows value migrating precisely toward the human capacities machines cannot yet supply, not evaporating along with the automated task.

• Acemoglu’s macroeconomics locates the real gains in reallocation toward human-complementary work, not in whatever remains of the automated task itself.

• The agentic-economy research shows intelligence itself becoming a cheap, transactable commodity, delivered on demand by agents rather than hoarded by institutions.

• Deming’s evidence shows a rising wage premium on distinctly human, non-routine skills precisely as routine cognition keeps getting automated out from under them.

We will:

• Re-found the curriculum on the appreciating human faculties, not the depreciating stores of memorised knowledge.

• Teach students to direct intelligence — to be the scarce ingredient — rather than compete uselessly with the abundant one.

• Treat “the machine can now do this” as the signal to move the learning goal up a level, never as the cue to drill the automated skill harder.

• Make the human comparative advantage — judgement, taste, agency, meaning — the explicit, named product the school exists to deliver.

• Start every curricular decision from the economics, not the tradition: value flows to scarcity, and the human is now the scarce thing.

2. The School We Have Optimises for the Losing Side of the Race

On the eve of the automobile, we doubled the budget for teaching everyone to shoe horses — and graded them, with real rigour, on the neatness of the nails.

Metaphor: A transmission engine built for one century is still running at full capacity in the next, faithfully delivering a cargo nobody downstream wants any more.

We hold:

• The modern school was engineered to mass-produce a specific kind of human being.

• Literate, numerate, punctual, compliant, able to store and reproduce information on demand.

• That human was the triumph of the industrial age and the exact match for its economy.

• We reject the comforting idea that this design is timeless rather than historically specific.

• We hold that it is, instead, precisely the human the machines have just made redundant.

• The most-rewarded act in most classrooms today is one a free chatbot performs flawlessly, in seconds, for nothing.

Why it holds:

• The historical evidence — the NBER histories, the industrial-model critiques — shows the standard school was built for a labour market that has since been replaced, not preserved.

• The OpenAI and ILO exposure studies show it is exactly the routine, procedural cognition the school’s core output that is most exposed to automation, not incidentally but by definition.

• Persistence-of-the-factory-model research (Yong Zhao) shows how little the institution itself has changed even as the world around it transformed completely.

• Assessment surveys already find students using AI to complete the very tasks schools grade — the proxy schools relied on to certify competence has quietly broken.

• The WEF’s own churn figures show the jobs the old curriculum targets are exactly the jobs disappearing fastest, not the ones holding steady.

We will:

• Audit every item on the timetable against one question: are we still teaching something the machines have already made worthless?

• Stop rewarding flawless reproduction of an automatable procedure as though it were the summit of academic achievement.

• Redirect the resources this frees — the hours, the teachers, the assessment budget — toward the faculties that are actually appreciating.

• Name the horse-shoeing honestly and in public, so the institution can no longer defend it to itself as rigour.

• Treat the mismatch between what we teach and what the world now needs as the emergency it is, not a topic deferred to a future working group.

3. Education’s Purpose Is to Build a Sovereign, Not to Fill a Vessel

You do not prepare someone to captain a ship by filling their head with the ocean; you build in them the judgement to read any sea they are ever likely to meet.

Metaphor: The old classroom is a container being topped up; the one we are demanding is a forge.

We hold:

• The old metaphor of education was the vessel: a mind to be filled with content, gradually, obediently.

• That metaphor made real sense when content itself was the scarce and valuable thing to hold.

• We reject it now, because the vessel’s contents have become free, infinite, and instantly retrievable.

• We hold the true metaphor to be construction: a person built, faculty by deliberate faculty.

• The product of school is not a stocked mind. It is a sovereign — one who owns their own reality.

• Education stops being transmission the moment it becomes the deliberate formation of a person.

Why it holds:

• Cognitive science — Willingham; Kirschner, Sweller and Clark — shows the faculties that matter are built through effortful use, not poured passively into a waiting mind.

• The OECD’s Learning Compass 2030 already reframes education around competencies and agency, not around content transmission for its own sake.

• Self-determination and agency research (Ryan and Deci; the OECD) shows sovereignty over one’s own learning is both buildable and, in practice, decisive.

• Character-education and purpose research (Damon; the Jubilee Centre) treats the formation of a person as a serious, teachable aim rather than a soft aspiration.

• The complementarity economics shows the built person, not the filled one, is what the new labour market and the new civic order both reward.

We will:

• Redefine the deliverable of school, in writing and out loud, as a person built, never as a syllabus merely covered.

• Organise the curriculum around named faculties to construct, using knowledge as the material, not as the final goal.

• Make agency, judgement and self-authorship explicit, assessed aims, not implicit hopes left to chance.

• Measure whether students can act, decide and create, not only whether they can recall on command.

• Hold the whole institution, every year, to the sovereign test: is this building an owner, or is it still filling a vessel?

4. What Is Automated Is Devalued; What Is Human Appreciates

Every time the tide of automation rises and covers another skill, it leaves the higher ground more valuable than the flood ever found it.

Metaphor: There is a shoreline moving through the economy right now, and everything below it is drowning in cheapness while everything above it climbs in price.

We hold:

• There is a law running underneath the whole of this transformation, whether or not a curriculum acknowledges it.

• Whatever machines can do becomes abundant, and abundance reliably destroys the economic value of a skill.

• Whatever remains distinctly human becomes, by the same motion, more scarce and more prized.

• We reject any curriculum that cannot say, item by item, which side of that law each subject sits on.

• We hold that the law tells you exactly what to teach: the faculties standing on the rising ground.

• Judgement, taste, originality, agency, relationship and meaning appreciate; storage and reproduction depreciate to zero.

Why it holds:

• Autor’s complementarity principle: automating a routine task raises, rather than lowers, the value of the human tasks that complement it.

• Brynjolfsson’s “Turing Trap” warns explicitly that imitating human output destroys economic value while augmenting human capability creates it.

• Deming’s data show a concrete, rising wage premium on social and non-routine human skills as routine cognition keeps getting automated.

• The WEF’s skills taxonomy shows analytical thinking, creativity, resilience and curiosity as the fastest-rising skills in the entire dataset.

• The offloading and cognitive-debt research shows the automatable faculties atrophy the moment they are handed to a machine — direct confirmation they were never really the point.

We will:

• Reorganise the entire curriculum around the appreciating faculties named across these sixteen theses and the sixteen subjects that follow from them.

• Apply one hard test to every subject on the timetable: does this appreciate or depreciate as the machines keep advancing?

• Move the learning goal up the value chain the instant a capability underneath it gets automated, rather than drilling it harder out of habit.

• Teach the law itself, explicitly, so students can navigate their own lifelong learning by it long after they leave us.

• Invest deliberately where the ground is rising — judgement, taste, agency, meaning — and abandon, without nostalgia, the flood plain beneath it.

5. AI Is the Greatest Tutor and the Greatest Cognitive Off-Switch at Once

The same fire that cooks the meal burns the house down; the sane response was never to ban fire, it was to learn to build a hearth.

Metaphor: Hand a child a hearth and they cook for a lifetime; hand them an open flame in an empty room and you have handed them a different future entirely.

We hold:

• The same technology produces opposite outcomes depending on nothing but how it is used.

• Used one way, it delivers some of the largest learning gains ever formally measured.

• Used another way, it hollows out the mind and leaves behind a real, measurable cognitive debt.

• We reject the framing that treats the technology itself as the variable that decides the outcome.

• We hold that the variable is never the tool. It is always how the tool is wired into the day.

• Neither banning it nor surrendering to it is an answer; the whole task, without exception, is design.

Why it holds:

• The MIT Media Lab EEG study shows LLM-assisted work measurably lowering brain engagement in real time — the off-switch, recorded on a scan, not asserted.

• Harvard and World Bank randomised trials show purpose-built tutors producing some of the largest learning gains ever recorded in the literature — the tutor, equally real.

• Wharton’s experiment shows the same tool harming or helping outcomes depending purely on whether it was guardrailed during practice.

• Stanford’s Tutor CoPilot research shows design that augments the human teacher produces real gains; raw, unguided deployment does not.

• The consistent lesson across the whole library is that outcome tracks configuration, never “AI” treated as an abstract, undifferentiated force.

We will:

• Reject both techno-surrender and moral panic as postures, and commit instead to the harder work of actual design.

• Mandate guardrailed, pedagogy-first tools that coach a student toward an answer rather than simply completing the task for them.

• Sequence every use of AI to amplify effort, never to substitute for the work that actually builds the mind.

• Evaluate every tool we adopt on learning outcomes measured with the tool removed, not on outcomes measured while it is still switched on.

• Treat design as the entire game worth playing here — the hearth, never the ban, and never the blank cheque either.

6. Struggle Is Not the Obstacle to Learning; Struggle Is the Learning

The muscle grows in the strain of the last repetition, never in the ease of setting the weight back down.

Metaphor: Every shortcut around a hard problem is a small loan taken out against a mind that has not yet been built to repay it.

We hold:

• There is a deep, decent, entirely understandable instinct to treat difficulty as a problem to remove.

• AI is the most powerful difficulty-removal machine ever placed in a classroom, and that is exactly its danger.

• We reject the idea that the effortful, uncomfortable part of a task is merely a tax on learning.

• We hold that the effortful part is not a tax at all. It is the learning, in its entirety.

• Remove the struggle and you remove the growth, leaving behind only a finished-looking output.

• Any school serious about learning has to protect productive struggle from the machine, on purpose, every day.

Why it holds:

• The Bjorks’ “desirable difficulties” research shows that making learning harder in the moment is exactly what makes it durable afterward.

• The MIT cognitive-debt study shows that when a machine performs the effortful work, the brain underneath simply fails to engage or grow.

• Wharton’s finding that practising with unguarded AI lowered later unaided scores is productive struggle removed and its price already being paid.

• Deliberate-practice research (Ericsson) locates the development of real expertise precisely in sustained effort at the edge of current ability.

• Retrieval-practice research (Roediger and Karpicke) shows the effortful act of recall, not easy re-exposure to the material, is what actually builds durable knowledge.

We will:

• Enforce struggle-first sequencing: the student attempts the hard thing unaided first, and only then brings AI in to check and extend it.

• Design tasks where the effort itself is the point, and reward the effort visibly, not merely the polish of the final output.

• Use AI to increase desirable difficulty deliberately — harder problems, tougher critiques — rather than to quietly remove it.

• Teach the concept of productive struggle to students directly, so they stop mistaking ease in the moment for progress over time.

• Protect the difficult, formative work from automation precisely where the temptation to automate it is strongest, not only where it is safe to.

7. Attention Is the Master Resource

Attention is the aperture through which the whole world enters a mind; narrow it to a flicker and nothing deep can ever get in again.

Metaphor: Every other faculty in this manifesto is a room in a house that attention alone has to build the door for.

We hold:

• Every faculty named in these sixteen theses runs, underneath, on one prior capacity: attention.

• Without sustained focus there is no deep learning, no real synthesis, and no genuine creation.

• Attention is also, right now, under deliberate, well-funded, professionally engineered assault.

• We reject the fatalism that treats this assault as an unavoidable feature of modern life.

• We hold attention to be a trainable faculty — a muscle that can be deliberately built or quietly allowed to waste.

• Whoever owns their attention owns everything built downstream of it; whoever loses it owns nothing, however much they know.

Why it holds:

• The neuroscience of attention shows focus is run by specific, trainable, genuinely fatigable brain networks, not by an innate and fixed personal trait.

• Mrazek and colleagues show brief, structured focus training measurably improves both working memory and test performance in real classrooms.

• Ward’s “brain drain” study shows the mere physical presence of a phone degrades cognitive capacity, even switched off, even face-down.

• The attention-economy literature — James Williams, the dark-patterns research, the OECD — documents an entire machinery engineered to capture exactly this resource.

• PISA data link classroom distraction directly to materially lower attainment, across countries and across subjects, not as a footnote but as a headline finding.

We will:

• Teach attention explicitly, as the first and most important subject on the timetable — trained, practised and formally measured, not assumed.

• Engineer the physical and digital environment for focus: phone-free defaults, single-purpose devices, and learning blocks genuinely free of notifications.

• Teach the attention economy itself as an adversary, by name, so students learn to defend their focus knowingly rather than by accident.

• Treat the ability to concentrate as the master competence every other competence in this manifesto quietly depends on.

• Frame it, honestly, as power: the person who can still concentrate will out-think everyone around them who no longer can.

8. Knowledge Is Not Obsolete; It Is the Precondition for Everything AI Makes Valuable

You cannot connect dots you do not have; “just look it up” hands a student a screen full of dots and not one working line between them.

Metaphor: A search engine is a warehouse with every part in stock and no one on the floor who knows how an engine goes together.

We hold:

• The most seductive error of this entire age is the claim that facts no longer matter, because everything is retrievable.

• We reject that claim, flatly, because cognitive science says it runs exactly backwards.

• Thinking, judgement and creativity all run on knowledge held in the mind, not on knowledge merely available on tap.

• We hold that a person cannot discern, synthesise, create or verify anything starting from an empty head.

• The more the machine can retrieve on demand, the more valuable becomes the human who already, genuinely knows.

• Knowledge is not the opposite of the appreciating faculties in this manifesto. It is their foundation.

Why it holds:

• Willingham shows critical thinking is domain-specific: people reason well only about material they already understand deeply, not about material they can merely locate.

• Kirschner, Sweller and Clark show reasoning runs in working memory, drawing constantly on a rich store of long-term knowledge, not on an external search bar.

• The National Research Council shows transferable competencies develop through rich content, and not, as is often claimed, instead of it.

• The offloading research shows externalised knowledge builds none of the mental schemas comprehension actually requires to function.

• Every appreciating faculty named across this manifesto — discernment, taste, systems sight — is shown, in the literature, to rest on deep prior knowledge, not to bypass it.

We will:

• Keep building deep, structured domain knowledge deliberately — AI raises its value, it does not remove the need to hold it.

• Reject the false binary of “skills, not facts” outright, and teach powerful knowledge and the faculties it makes possible, together, on purpose.

• Use AI to deepen a student’s knowledge, and refuse, as a matter of policy, to let it excuse the absence of any.

• Build the schema first, every time, and let retrieval extend it afterward — never let retrieval stand in for the schema itself.

• Treat a knowledge-rich curriculum as the precondition for every single faculty named in this manifesto, not as a competing priority to trade against them.

9. The Scarce Skills Are Judgement, Taste, Agency and Meaning

We are grading children on the size of their memory in an age of infinite memory, and quietly ignoring the four things now beyond price.

Metaphor: A library card was once worth a fortune; today the four things a machine still cannot hold are worth the fortune instead.

We hold:

• Ask what a machine genuinely cannot do, and the honest answer names the new curriculum outright.

• It cannot decide what is worth doing in the first place. That is judgement, and it is irreducibly human.

• It cannot reliably tell the excellent from the merely plausible. That is taste, and it too is human.

• We reject any curriculum that leaves judgement and taste to chance rather than teaching them directly.

• We hold that agency — the ability to want, to initiate, to own an outcome — belongs on the same list.

• Meaning belongs there too: a machine can never answer, on a student’s behalf, why any of it finally matters.

Why it holds:

• The complementarity literature (Autor; Brynjolfsson) locates enduring economic value precisely in judgement and in the human residual left after automation.

• The Frontiers work on AI dependence shows taste and creativity measurably erode when they are routinely offloaded — direct confirmation they are scarce, human, and fragile.

• Agency research (Ryan and Deci; the OECD) shows initiative and a sense of ownership over one’s own work are both decisive and genuinely buildable.

• Damon’s and Seligman’s work shows meaning and purpose are powerful, teachable and protective, not simply the byproducts of a well-run curriculum elsewhere.

• The WEF’s rising-skills data — creative and analytical thinking, resilience, curiosity — map with real precision onto exactly these four faculties.

We will:

• Make judgement, taste, agency and meaning explicit, assessed aims of the curriculum, named on the same page as literacy and numeracy.

• Assess “which is best, and why,” “what would you actually do,” and “why does this matter” — not only “what is the correct answer.”

• Redesign assessment, structurally, to reward these four scarce faculties, because what we choose to test is what we end up building.

• Stop measuring the size of a student’s memory and start measuring, deliberately, the quality of their judgement instead.

• Treat these four as the core of the timetable, not its enrichment — the centre of the school day, never its margin.

10. Teach With AI and Defend the Mind From It — Both, at Once

A soldier is trained both to use the weapon and to survive it; teaching only one of the two produces, reliably, a casualty.

Metaphor: One mandate without the other is half a shield, carried proudly into a fight it was never built to survive.

We hold:

• There are two mandates here, and they have to run in the same room, on the same day, without exception.

• The first: teach real fluency with AI, because the economy now demands it without any real room for negotiation.

• The second: defend the mind from AI, because misused, it quietly erodes the very faculties this manifesto exists to build.

• We reject a school that runs only the first mandate — it produces confident incompetents who collapse the moment the tool is taken away.

• We reject, equally, a school that runs only the second — it produces disciplined minds that are simply unemployable.

• We hold that the entire art here is holding both mandates at once, deliberately, every single day, without letting either one drift.

Why it holds:

• The future-of-work evidence — the WEF, the exposure studies — makes AI fluency effectively non-negotiable: the first mandate, confirmed from the labour-market side.

• The cognitive-debt, offloading and over-reliance research makes defending the mind equally non-negotiable: the second mandate, confirmed from the cognitive-science side.

• The Nigeria and Harvard studies show, concretely, what real fluency delivers when it is taught well and taught on purpose.

• The MIT and Wharton findings show, just as concretely, what happens the moment the mind is left undefended.

• Read together, the two literatures prove that neither mandate alone is survivable — this is not a matter of balance, it is a matter of arithmetic.

We will:

• Run both mandates deliberately and simultaneously: fluency and defence, inside the same curriculum, on the same syllabus.

• Teach students to use AI with real power, and, in the same breath, to protect what has to stay in their own heads regardless.

• Distinguish clearly, subject by subject, between tasks that are safe to delegate and faculties that must still be built entirely unaided.

• Refuse the false choice between adoption and protection outright — the honest answer is both, designed together, not traded against each other.

• Judge every school, including our own, on whether it produces people who are fluent with the machine and sovereign without it.

11. The Goal Is the Owner, Not the Employee

For a century we trained people to be excellent passengers; the age of abundance pays out only to those who can actually drive.

Metaphor: A hundred years of polishing the passenger seat, in an age that is about to reward whoever can reach the wheel.

We hold:

• The industrial school was designed, quite deliberately, to produce a reliably good employee.

• Reliable, compliant, able to follow instructions along a predetermined and legible path.

• That path is dissolving in real time, and with it the market value of the passenger.

• We reject the idea that passivity, however well-mannered, remains a viable strategy for a young person.

• We hold that the new economy rewards the owner instead — of businesses, of assets, of their own direction.

• It rewards the protagonist who acts, not the extra who waits, however patiently, to be told where to stand.

Why it holds:

• Agency research (the OECD; Ryan and Deci) shows the self-authoring disposition is both decisive in outcomes and genuinely buildable through instruction.

• Entrepreneurship evidence (the OECD; Amit and Zott) shows value-creation and ownership are teachable capacities, not fixed personality traits some students simply lack.

• The ownership-divide economics shows returns concentrating steadily in owners as more and more of routine labour gets automated away.

• The collapse of fixed, linear career paths removes the very structure that once made a passive strategy viable for an ordinary student.

• Duckworth’s grit research ties real achievement to the self-driven pursuit of one’s own chosen goals, not to compliant pursuit of someone else’s.

We will:

• Flip the hidden curriculum from compliance to ownership, on purpose, and reward the student who initiates and owns an outcome.

• Teach value creation, capital and agency as core subjects, so students can learn to own something, not only to earn a wage from it.

• Give students real responsibility with real stakes, so ownership is genuinely practised in school, not merely preached at them from the front.

• Frame the aim of school explicitly and out loud: we are building owners of their own reality, not passengers riding in someone else’s.

• Treat passivity, wherever we find it, as the failure mode it now is, and treat initiative as the achievement it has always deserved to be.

12. Money, Capital and Ownership Must Be Taught to Everyone

We teach children the water cycle in exhaustive detail and the money cycle not at all — and then we act surprised at how many of them drown.

Metaphor: A whole curriculum built around one cycle nature runs for free, and total silence about the one cycle a person actually has to run themselves.

We hold:

• The single largest omission in the standard curriculum, measured against its consequences, is money.

• Not budgeting tips at the margins, but the real mechanics of capital, ownership and wealth.

• We reject the polite fiction that this omission is neutral, an accident of a crowded timetable.

• We hold instead that it quietly serves those who already hold capital and abandons everyone who does not.

• The difference between earning wages and owning assets that earn on your behalf is not a footnote.

• In an economy where returns concentrate steadily in owners, teaching this to everyone is a matter of justice, not enrichment.

Why it holds:

• Financial-literacy research (Kaiser and Lusardi; Hastings et al.) links financial capability causally to real wealth and real long-run outcomes, not merely to confidence.

• OECD PISA data show large, deeply unequal gaps in young people’s financial literacy — measurable, documented, and left almost entirely unaddressed.

• The macroeconomics of AI shows returns shifting steadily from labour toward capital, which widens exactly this divide rather than narrowing it.

• Winner-take-most dynamics concentrate gains, disproportionately, among the owners of systems rather than among the people operating inside them.

• The GFLEC and World Bank evidence shows financial capability can be built deliberately, and built especially effectively when it starts early.

We will:

• Teach money, capital and ownership universally and early — the real mechanics of wealth, for every student, not a self-selecting few.

• Teach the earning-versus-owning distinction explicitly, along with the real pathways that connect one to the other.

• Connect financial literacy to entrepreneurship directly: building and owning value, not simply selling one’s own labour by the hour.

• Use concrete, real practice wherever possible — equity, compounding, allocation, risk — rather than abstract lessons about money in general.

• Treat financial illiteracy as the equity emergency it already is, and treat its remedy, correctly, as empowerment rather than charity.

13. Meaning Is the Emergency of an Age of Abundance

When the desert finally floods, people do not die of thirst any longer. They drown instead. Abundance has its own, very different way of killing.

Metaphor: Remove the old scaffolding of necessity and a person either builds a new structure to stand on or falls straight through the floor.

We hold:

• For most of history, survival and work supplied a person’s meaning by sheer necessity, without being asked to.

• As machines do more and abundance grows, that external scaffolding is quietly falling away beneath us.

• When work becomes optional or fluid, the question “why do anything at all?” arrives with real force.

• We reject the idea that this question can safely be left to chance, or to religion, or to the market.

• We hold that answered badly, it produces nihilism, drift and despair, and answered well, it produces resilience.

• Meaning is not a luxury item on the curriculum. It is the emergency sitting quietly at its exact centre.

Why it holds:

• Damon’s research shows purpose is a powerful driver of engagement, resilience and wellbeing specifically in young people, not only in adults.

• The Frontiers adolescent research links a felt purpose in life directly to protection against depression, not merely to reported happiness.

• Seligman and Adler show wellbeing and meaning can be deliberately cultivated inside a school, using methods that already exist and already work.

• Rising adolescent mental-health strain, amplified by the attention economy (CIGI), makes the need for this urgent rather than aspirational.

• Self-determination theory ties durable motivation directly to purpose — an internal source of drive that survives automation where external rewards do not.

We will:

• Put meaning, ethics and the examined life at the centre of the curriculum, and teach them with the same seriousness as mathematics.

• Help every student build a source of meaning that is independent of any single job, title or external validation.

• Teach the great questions directly — what makes a good life, what is actually worth wanting — as core material, not as an elective.

• Anchor real learning in contribution and purpose, so that student effort feels meaningful rather than merely compliant.

• Treat the capacity to generate one’s own meaning as a survival skill for an age of abundance, and teach it accordingly.

14. The Human Must Hold the Reins

However perfect the autopilot, a court still asks for the name of the captain — because accountability is a thing only a human being can carry.

Metaphor: A machine can steer the ship all night; only a person can be woken up and asked to answer for where it ends up.

We hold:

• Machines can decide, and machines can act, but a machine cannot be responsible for what follows.

• Responsibility — moral responsibility, legal responsibility — is an irreducibly human property, not a feature to be engineered.

• As agents do more and more in the world, the question of who owns the outcome only sharpens further.

• We reject any design that quietly lets accountability diffuse into the system rather than resting on a named person.

• We hold that every consequential automated decision needs a human being who holds the reins on it.

• This demands the judgement to know precisely when to trust the machine and when, deliberately, to override it.

Why it holds:

• The WEF’s AI-governance work stresses accountability directly, and insists on keeping humans answerable for what their agents actually do.

• Hadfield and Koh show agents require human institutions and real accountability structures around them in order to function safely at all.

• The US Department of Education insists explicitly on humans staying in the loop and remaining accountable for decisions made about students.

• Human-AI interaction research (Amershi et al.; Bansal et al.) shows appropriate reliance on a machine is a hard, genuinely learnable skill, not an instinct.

• The hallucination and verification evidence makes human judgement, in practice, the last real line of defence in any automated system.

We will:

• Teach the ethics and the practice of delegation explicitly: when to trust a machine, when to override it, and who is accountable either way.

• Build the judgement to recognise a machine’s limits, and to own the outcomes of its actions as though they were our own.

• Make moral reasoning about automated systems a core, examined subject, not an afterthought bolted onto a computing class.

• Practise accountable decision-making with AI genuinely in the loop, with a human owning the result at every step.

• Frame responsibility, explicitly, as a distinctly human role that grows in importance exactly as fast as the machines do.

15. Every One of These Faculties Is Trainable

We stand in front of a locked door holding the key in our own hand, insisting the door cannot be opened — because opening it has never been our habit.

Metaphor: The evidence is not the missing piece. The habit of ignoring the evidence is the missing piece.

We hold:

• The reflex objection to this entire manifesto is that its aims are simply unteachable in a normal school.

• The claim runs that attention, judgement, agency, taste and meaning are gifts, not skills that can be built.

• We reject that claim outright. The evidence says otherwise, decisively, across every one of these sixteen faculties.

• We hold that each one rests on a real research literature showing it is measurable, predictive, and buildable on purpose.

• The barrier to a school for owners has never once, in this evidence base, been a matter of feasibility.

• The barrier is habit. We keep teaching the horse-shoeing because it is what we have always, comfortably, done.

Why it holds:

• Attention training works — Mrazek, Jha — and focus is demonstrably buildable through structured, repeatable practice, not fixed at birth.

• Metacognition and self-regulation are, per the EEF and Hattie, among the single most teachable, highest-impact outcomes measured in education research.

• Financial capability, social-emotional skills, character and entrepreneurship all carry evidence bases showing schools can build them directly and reliably.

• Project-based, mastery and apprenticeship models show the appreciating faculties develop reliably under the right structural conditions, not by accident.

• The 179-document library behind this whole series is, in aggregate, the proof itself: these are trainable capacities, not fixed personal traits.

We will:

• Stop treating the appreciating faculties as innate gifts, and start teaching them with the same seriousness the system already gives to literacy.

• Use the evidence cited across this entire series to design real, deliberate instruction for each faculty, one at a time, on purpose.

• Reallocate the resources freed from teaching the automatable material toward building the faculties that are actually trainable and actually scarce.

• Pilot, measure and scale what demonstrably works, faculty by faculty, rather than waiting for a single grand redesign that never quite arrives.

• Name the real barrier honestly — habit, not impossibility — and overcome it, because the key has been sitting in our hand the whole time.

16. Build Sovereigns — or Hand the Next Generation a World They Can Only Serve

We are about to hand a generation the most powerful tools ever forged; the only real question is whether we hand them over as masters or as servants.

Metaphor: Two children stand beside the same extraordinary machine. One was taught to direct it. One was taught only to be graded faster by it. The machine cannot tell them apart. The world will.

We hold:

• This is the thesis every one of the other fifteen has been quietly building toward, from the very first page.

• The machines are going to be extraordinary. That much, at this point, is no longer in serious dispute.

• What is not certain is whether the humans standing beside them will be made extraordinary too.

• We reject the neutral-sounding option of simply waiting to see how this resolves itself over time.

• We hold that an education built to form sovereigns produces people who own and direct the machines beside them.

• An education that clings to the old timetable produces people the machines go on, quietly, to replace. There is no third door.

Why it holds:

• The whole of the future-of-work evidence shows a divide opening, in real time, between those who direct AI and those it displaces.

• The ownership-divide economics shows the identical split appearing in capital: owners rise, and the unprepared are left fully exposed.

• The cognitive-debt and offloading research shows precisely what happens to a mind handed to a machine without any defence at all.

• The complementarity and agency literatures show the sovereign path is real, evidenced and buildable, not merely a hopeful rhetorical flourish.

• History shows education systems can transform when the necessity is finally, honestly faced — and that the cost of delay is always paid by children, never by the adults who chose to wait.

We will:

• Choose, deliberately and now, to build sovereigns — because that has been the whole point of this exercise from thesis one.

• Enact the shift in full: burn the old timetable, build the sixteen faculties, run both mandates, every day, without exception.

• Lead the change through teachers and evidence, urgently, treating it as the emergency it plainly is, not a slow-moving reform.

• Give every child, not only the already advantaged, the education of an owner rather than the education of a passenger.

• Accept the assignment in full: the next generation inherits the machines, or is inherited by them, and that outcome is ours to decide, starting now.

We Declare

Strip these sixteen theses to a single sentence and it is this: when intelligence becomes abundant, the purpose of education must change from filling minds with knowledge to building humans who are sovereign over their own. Everything else — the twenty-four challenges named in the first article of this series, the fourteen forces named in the second, the sixteen subjects named in the third — is the working-out of that one change, restated sixteen different ways so that no reader can mistake it for a slogan rather than a programme.

It is worth being honest about the size of what is being asked here. This is not a new module, a pilot programme, or a policy tweak scheduled for review in three years. It is a different answer to the oldest question an education system can be asked: what is a person for? The industrial age answered “to be a productive worker,” and built a school, brick by brick, to match that answer exactly. The age of abundant intelligence demands a different answer — “to be a sovereign, an owner of their own attention, judgement, capital and meaning” — and it will demand, in turn, a different school built to match it. We are the generation with the standing, the evidence and the narrow window of time to build it, before another cohort of children passes all the way through the old machine and out the other side, unprepared, on our watch.

The comfortable move, as always, is to wait. Run one more pilot. Convene one more panel. Let the institution change, as it always has, at its own accustomed and unhurried pace. But the children sitting in a classroom this year do not have that pace available to them. They graduate into the transformed world within a decade, and the education they receive is the one we choose to give them now, inside a system still optimised for a world that has already, provably, ended. To move slowly here is not caution, whatever it calls itself in a committee meeting. It is a decision, made on a child’s behalf and entirely without their consent, to prepare them for the losing side of a race we can already see the shape of.

Therefore, we — researchers who built the 179-document evidence base beneath these sixteen theses, teachers who will have to teach differently starting Monday, parents who will not get a second attempt at this childhood, and, above every other stakeholder in this argument, the students this system is about to either build or discard — do hereby declare that the purpose of education has changed, whether or not the institution has yet caught up with the fact. We do not ask permission for this declaration, because self-evident truths are not, by their nature, subject to a vote. We ask only that every reader who has followed these sixteen theses this far now go and act as though they were true, because they are.

We will not wait for a national curriculum to move first. We will not wait for a testing regime built for the twentieth century to bless a twenty-first-century child. We will not wait for consensus among people with no child currently inside the system paying its cost. We will build the school for owners in the rooms we already control — one classroom, one department, one institution at a time — and we will measure it honestly against the sovereign test laid out across these sixteen theses, and we will publish what we find, whether it flatters us or not.

The machines are going to be extraordinary. That was never actually in question. Make the humans extraordinary too. Build sovereigns — people who own their attention, their judgement, their capital, their meaning and their choices — and hand the next generation a world they can master rather than one they can only serve. That is the assignment. It has always been the whole of the assignment. And the clock on it, whatever the timetable says, is already running.

Built on a purpose-built ENSI research library of 179 primary documents across 30 research angles — from the neuroscience of attention and the science of learning to the economics of the agentic labour market and the psychology of meaning — spanning the OECD, UNESCO, the WEF, the World Bank, NBER, the MIT Media Lab, Stanford, the EEF, CASEL and dozens more. Part four, and the capstone, of the ENSI “Education for the Agentic Age” series.

This article does something the previous two only pointed at. It burns the timetable. Not reforms it, not augments it, not bolts an “AI literacy” module onto the side of it — replaces the organising principle entirely. The old principle was content: a curriculum is a list of bodies of knowledge to be transmitted and tested. That principle made sense when knowledge was scarce and locked in the heads of experts and the pages of expensive books. It is now catastrophic, because knowledge is abundant and nearly free, and a curriculum built to transmit it is optimising for the one thing the machines have made worthless.

The replacement principle is faculty. A curriculum should be a list not of things to know but of human capacities to build — specifically, the capacities that appreciate rather than depreciate as intelligence becomes abundant. The second article traced fourteen forces reshaping work and society and showed they converge on a small set of these appreciating faculties. This article takes that set — sixteen of them — and defines each one properly: what it actually is, the evidence that it is real and buildable, and how a school would go about constructing it in a young human.

A warning before we start, because the names are deliberately provocative. These are not the sanitised, committee-approved labels of a national curriculum. They are named to be remembered and to be honest about their ambition: to build sovereigns, not subjects — people who own their attention, their judgement, their capital and their choices, rather than people who are managed by the systems around them. “Cognitive Sovereignty” is a bolder claim than “study skills,” and it is meant to be, because the stakes are bolder. This is not a gentle update. It is a different theory of what a person is for.

The sixteen fall into four clusters, and the clusters matter because they describe a whole person. The Sovereign Mind is the set of inner faculties that run on your own neural hardware and that no machine can rent from you: trained attention, disciplined discernment, the ability to see whole systems, and the imagination to build futures. The Builder is the set that turns intelligence into value in the real world: creating value, commanding fleets of machines, exercising taste, and mastering physical matter. The Sovereign Self is the operating system of a free person: agency, health, capital, and meaning. And The Arena is the set for acting among other people and other machines: moving people, leading them, learning under real stakes, and holding the reins of the systems you command.

Notice what each one buries. Cognitive Sovereignty and Discernment absorb and replace the fragments of “media literacy” and “study skills” that were never taught seriously. Systems Sight and Building Futures do the real work that “geography” and “history-as-dates” were pretending to do — understanding how the world’s systems interconnect and how they change over time. Commanding the Swarm replaces “computer studies.” The Value Engine and Capital & Power fill the gaping hole where money, ownership and value creation should always have been. Nothing worth knowing is thrown away — deep knowledge is the precondition for every one of these faculties, as the cognitive science insists — but knowledge stops being the point and becomes the raw material from which the faculties are built.

One more thing must be said plainly, because it is the load-bearing claim under all sixteen. None of these faculties is soft, vague or unteachable. Each is backed by a real research literature showing it is a genuine capacity, that it predicts real outcomes, and that it can be deliberately developed. Attention is trainable. Judgement is trainable. Agency, taste, financial capability, persuasion, metacognition — every one has an evidence base, cited below, showing schools can build it. The reason we don’t is not that we can’t. It is that we have been busy teaching the horse-shoeing.

So here are the sixteen subjects of a school for owners — the curriculum you would design if you started from the person you were trying to build, in the world they are actually going to live in, rather than from the timetable you inherited from a world that has already ended.

The sixteen subjects, in short

1. Cognitive Sovereignty — trained, defended attention; the master resource of the century and the thing every other faculty runs on.

2. Discernment — knowing what is true when infinite confident content is free; sourcing, calibration, disciplined doubt.

3. Systems Sight — seeing wholes, feedback loops and second-order effects; the synthesis machines are worst at.

4. Building Futures — foresight and world-building; imagining what doesn’t exist yet and reverse-engineering the path to it.

5. The Value Engine — how value is really created and captured; the world from the builder’s chair, not the consumer’s.

6. Commanding the Swarm — directing fleets of AI agents: specify, delegate, verify, judge; the new literacy.

7. Taste & Origination — generating genuinely new ideas and the judgement to know which are good; the moat when content is infinite.

8. Mastery of Matter — making real things in the physical world; keeping a species’ grip on reality.

9. Living as an Owner — agency, initiative, self-authorship; being the protagonist rather than the NPC.

10. The Body as Base — vitality as the substrate of cognition; the one machine you must maintain yourself.

11. Capital & Power — money, ownership, equity versus wages, allocation; the mechanics of wealth taught to everyone.

12. Why Anything At All — meaning, values, the examined life; the antidote to nihilism when work is optional.

13. The Power to Move People — rhetoric, story, persuasion, negotiation; human-to-human influence that never automates.

14. Command of People — trust, teams, leadership, emotional regulation; the relationship-dense core of the future economy.

15. Playing for Real — high-agency learning under genuine stakes; school as arena, not waiting room.

16. Who Holds the Reins — the human-machine contract: when to delegate, where a human must stay accountable.

Cluster I — The Sovereign Mind

1. Cognitive Sovereignty

Metaphor: In a world designed to steal your attention every three seconds, the ability to hold it is the difference between owning your mind and renting it out for free.

Definition: This is the disciplined command of your own attention.
It is the capacity to focus deeply, at will, on one hard thing.
It is the trained resistance to a technology built to fracture exactly that.
It replaces the assumption that focus is a personality trait you either have or lack.
It treats attention as a muscle: buildable, and, if neglected, wasteable.
It is first on the list because every other faculty here runs on it.

Why it holds:

• The neuroscience of attention (Petersen and Posner; Corbetta and Shulman) shows focus is run by specific, trainable, fatigable brain networks — a real capacity, not a mood.

• Mrazek and colleagues found that just two weeks of focus training reduced mind-wandering and raised both working memory and standardised-test performance.

• Jha and colleagues showed attention training measurably strengthens distinct sub-systems of attention — alerting, orienting, conflict-monitoring.

• Ward’s “brain drain” study shows the modern environment actively degrades this capacity, making its deliberate cultivation urgent.

• The flow-state research shows sustained deep focus is both a peak-performance state and a developable skill.

How to build with it:

• Teach attention explicitly and progressively: daily deep-work blocks, single-tasking as a norm, focus practised and measured like reading.

• Use attention and mindfulness training with an evidence base — short, regular, built into the day.

• Engineer the environment for sovereignty: phone-free defaults, single-purpose devices, notification-free learning cores.

• Teach students the attention economy as the adversary (subject 16 and Article 1) so they defend their focus knowingly.

• Frame it as power, not restriction: the person who can concentrate can out-think everyone who can’t.

2. Discernment

Metaphor: When every stranger is a fluent, confident, tireless liar, the person who can tell truth from performance is the only one who can be trusted to steer.

Definition: This is the disciplined pursuit of what is actually true.
It is knowing how to source, weigh, calibrate and doubt.
It is the reflex of checking a claim before acting on it — human or machine.
It replaces the fragmentary, unserious “media literacy” of the old curriculum.
It is lucidity: probabilistic thinking, evidence, and immunity to fluent nonsense.
In an ocean of free synthetic plausibility, it is the faculty that keeps you from drowning.

Why it holds:

• The detector studies (Weber-Wulff; the Stanford bias study) prove that surface fluency carries no signal of truth — so discernment cannot be automated and must be human.

• Microsoft Research found trust in AI displaces critical scrutiny, making trained skepticism a scarce corrective.

• The Elaboration Likelihood Model (Petty and Cacioppo) explains why fluent confidence persuades under low effort — the exact vulnerability discernment defends.

• The EEF and the metacognition literature show evaluation and self-monitoring are teachable, high-leverage skills.

• Willingham’s work shows critical evaluation is possible only atop real domain knowledge — grounding discernment in a knowledge-rich curriculum.

How to build with it:

• Make verification a trained reflex: every claim, especially a fluent AI one, gets sourced and checked before use.

• Teach the epistemic toolkit — evidence, probability, calibration, base rates, bias — as a core discipline.

• Assign adversarial work: find the error, argue the opposite, grade the model, rate your own confidence.

• Ground it in deep knowledge; you can only fact-check what you understand (subject buries nothing worth knowing — it uses it).

• Reward the student who catches the falsehood, not just the one who produces the fluent answer.

3. Systems Sight

Metaphor: Anyone can see the chess pieces; the master sees the forces between them — and that invisible board is where every real problem lives.

Definition: This is the capacity to see wholes, not just parts.
It is understanding feedback loops, incentives, second-order effects and emergence.
It is connecting distant domains into a single working model of how something behaves.
It replaces “geography” and “general studies” with the actual skill they gestured at: understanding how the world’s systems interlock.
It is the one cognitive move machines are worst at over long horizons and novel situations.
It is what turns a pile of facts into the ability to predict and intervene.

Why it holds:

• The complementarity literature (Autor; Brynjolfsson) shows value concentrating in the integrative human judgement machines cannot supply.

• Systems Sight is the synthesis layer above retrieval — and the offloading research shows retrieval alone builds none of it.

• The National Research Council’s Education for Life and Work identifies transfer and the integration of knowledge as the hard, high-value outcomes of deeper learning.

• Cognitive science (Willingham; Kirschner-Sweller-Clark) shows integrated understanding is built on rich, well-organised knowledge — systems thinking is knowledge compiled into models.

• The forces of Article 2 — task recombination, second-order economic effects — reward exactly those who can model whole systems.

How to build with it:

• Teach through interdisciplinary, systems-based problems, not siloed subjects — real questions that cross the old boundaries.

• Use maps, models, causal loops and simulations to make invisible structure visible.

• Assess the ability to predict second-order effects and to intervene in a system, not to recall its parts.

• Build the rich knowledge base first — you cannot see a system you know nothing about — then teach the integration.

• Reward students for connecting distant ideas into a working model, the signature move of Systems Sight.

4. Building Futures

Metaphor: The people who shape the world are not the ones who predict the weather but the ones who design the buildings the weather will hit — foresight is a construction skill, not a spectator sport.

Definition: This is the disciplined imagination of what does not yet exist.
It is scenario thinking, foresight, and the reverse-engineering of a path to a chosen future.
It is the refusal to treat the future as something that merely happens to you.
It replaces “history as a fixed list of the past” with the active use of time — learning from the past to design what comes next.
It is entrepreneurial and civic at once: the capacity to author futures, not just inhabit them.
It is ENSI’s home discipline, and it is teachable.

Why it holds:

• The foresight tradition (and ENSI’s own body of work) treats future-building as a capability institutions and individuals can develop, not a gift.

• The agency and self-determination research shows an internal, future-oriented locus of control drives achievement and wellbeing.

• Entrepreneurship research (the OECD’s entrepreneurial-mindset work) frames opportunity recognition and future-orientation as learnable dispositions.

• Building Futures is applied Systems Sight over time — grounded in the same evidence that integrated modelling is buildable.

• Damon’s purpose research links a compelling vision of one’s own future to motivation and resilience.

How to build with it:

• Teach scenario thinking and world-building — students imagine multiple futures and design paths toward the ones they want.

• Use history not as a dead list but as a laboratory for how change actually happens.

• Set projects that require inventing something that doesn’t exist yet and planning its path to reality.

• Connect it to entrepreneurship (subject 5): a future imagined is a value proposition waiting to be built.

• Cultivate the disposition that the future is authored, not awaited — the mindset of a builder, not a forecaster.

Cluster II — The Builder

5. The Value Engine

Metaphor: School has always taught children to look for the best chair in the room; this subject teaches them to build chairs other people will pay to sit in.

Definition: This is the study of how value is actually created and captured in the world.
It is seeing the economy from the builder’s chair, not the consumer’s or the employee’s.
It is the entrepreneurial mindset: spotting a need, building a solution, delivering it, capturing the return.
It replaces the total absence of value-creation from the curriculum — the hole where the most important economic skill should be.
It is not “business studies” as memorised definitions; it is making things people want.
As AI collapses the cost of building, this becomes a general necessity, not a niche.

Why it holds:

• The OECD’s entrepreneurship-education work establishes the entrepreneurial mindset — initiative, opportunity recognition, value creation — as a real, teachable capacity.

• Amit and Zott’s framework shows precisely how value is created and captured through novel configurations of resources — a formal, learnable model.

• The agentic-economy evidence shows AI radically lowering the cost of building, amplifying the leverage of individual value-creators.

• NBER work on business dynamism demonstrates how much economic vitality — and individual opportunity — depends on new value creation.

• Brynjolfsson, Li and Raymond, and Noy and Zhang, show AI most rewards those who direct it to create, not just to consume.

How to build with it:

• Have students actually build and ship things real people want — products, services, projects with real feedback.

• Teach how value is captured, not only created — pricing, business models, the mechanics of an offer.

• Cultivate the entrepreneurial dispositions: initiative, resilience, opportunity-spotting, comfort with risk.

• Use AI as the force-multiplier it is — students build far more ambitious things than a solo human once could.

• Make “you made something real that someone valued” the highest form of achievement in the school.

6. Commanding the Swarm

Metaphor: The future does not belong to the best violinist; it belongs to the conductor who can make a hundred machine-musicians play as one.

Definition: This is the literacy of the agentic age: directing fleets of AI agents.
It is specification, delegation, verification and the taste to judge machine output.
It is conducting intelligence, not memorising the syntax of any one tool.
It replaces “computer studies” — menus and button-locations — with command of thinking machines.
It is the difference between the person who directs the agents and the person the agents replace.
It is a genuinely new skill, and, like conducting, it is learnable and deep.

Why it holds:

• The agentic-economy research (Microsoft Research; Hadfield and Koh) describes an economy of agents that humans must learn to direct and govern.

• Prompt-pattern work (White et al.) shows orchestration is an engineerable, teachable skill with real structure, not a knack.

• Human-AI interaction research (Amershi et al.; Bansal et al.) shows that effective collaboration and appropriate reliance are hard, learnable competencies.

• Brynjolfsson, Li and Raymond found the largest gains flow to workers who effectively direct AI — orchestration has measurable economic return.

• The verification literature makes clear that commanding the swarm requires the judgement to check its work — tying it to Discernment (subject 2).

How to build with it:

• Teach specify-delegate-verify as an explicit skill: break down goals, assign them to agents, judge and integrate the results, own the outcome.

• Move students from producing answers themselves to directing and verifying systems that produce them.

• Practise real human-AI teaming on substantial projects, with the human accountable for quality and truth.

• Teach taste in output — recognising when the machine’s work is genuinely good — as inseparable from orchestration.

• Make AI fluency a baseline literacy for every student, as fundamental as reading and number.

7. Taste & Origination

Metaphor: When a machine can paint ten thousand competent pictures a second, the priceless human is the one who can make the single strange, true one — and know why it matters.

Definition: This is two joined faculties: making the genuinely new, and judging what is good.
Origination is the courage and capacity to produce something that did not exist before.
Taste is the refined judgement to tell the excellent from the merely plausible.
It replaces “art as decoration” with creativity as the central economic and human act.
Machines generate the average of everything; taste and origination are the human residual.
When plausible content is infinite and free, these are the moat.

Why it holds:

• Generative models regress to the mean of their training data, so genuine novelty and the judgement of quality are irreducibly human.

• The Frontiers study on AI dependence and creativity warns that offloading to the machine erodes the very origination that is now scarce.

• Expertise research (Ericsson) shows refined judgement — taste — is built through deliberate practice; it is trainable.

• The WEF ranks creative thinking among the fastest-rising skills as routine production automates.

• Brynjolfsson’s “Turing Trap” argues the economic prize lies in the distinctly human capacities — originality and judgement chief among them.

How to build with it:

• Teach taste through immersion in the excellent across domains, so the student’s internal standard rises.

• Protect un-assisted origination — the blank page — as a valuable ritual, then use AI to extend, not replace, the student’s ideas.

• Assess judgement directly: “which of these is best, and why,” making the student the editor and critic.

• Reward the surprising, the original and the well-judged over the fluent and the average.

• Teach students to generate before they generate-with, so their own creative voice forms first.

8. Mastery of Matter

Metaphor: A civilisation that can only manipulate symbols and forgets how to touch the physical world has handed away the steering wheel and kept the horn.

Definition: This is fluency in the physical, material world.
It is making real things — building, engineering, crafting, repairing.
It is the intuition for how matter behaves that only comes from working with it.
It replaces a purely abstract, screen-bound education with hands on real stuff.
It matters more, not less, as the digital world floods with the synthetic and the fake.
It keeps a person — and a species — anchored to reality that cannot be hallucinated.

Why it holds:

• Embodied-cognition research shows thinking itself is grounded in physical action and manipulation, not floating above it.

• The embodied-mathematics work (Nathan and Walkington) shows bodily action and gesture drive genuine conceptual insight.

• Cognitive-apprenticeship theory (Collins, Brown, Holum) — model, coach, scaffold, fade — comes from the crafts and remains a powerful learning model.

• Apprenticeship and dual-VET evidence (OECD; Cedefop) shows work-based, hands-on learning produces strong, durable competence and employment outcomes.

• As AI makes symbolic work abundant, the physical and material — which agents cannot directly touch — rises in relative human value.

How to build with it:

• Give students real making — workshops, labs, engineering, craft, building — as a core strand, not an elective.

• Use apprenticeship and cognitive-apprenticeship models: learn by doing alongside a skilled practitioner.

• Anchor abstract concepts in physical construction, where understanding is tested against reality.

• Teach the intuition for materials, tools and systems that only hands-on work builds.

• Value the ability to make and mend real things as a form of sovereignty over one’s physical world.

Cluster III — The Sovereign Self

9. Living as an Owner

Metaphor: In every story there is the protagonist who acts and the extra who waits to be told where to stand; school has spent a century producing extras.

Definition: This is the disposition of agency — being the author of your own life.
It is initiative: acting without waiting to be told, taking responsibility as default.
It is an internal locus of control — the belief and habit that you move the world, not only that it moves you.
It replaces the hidden curriculum of compliance with the hidden curriculum of ownership.
It is living as the owner of your situation, your knowledge, your reality — not a passenger in them.
It is the spine of every other faculty here; without it, the rest are inert.

Why it holds:

• Self-determination theory (Ryan and Deci) establishes autonomy as a root of motivation, persistence and wellbeing — agency is foundational, not optional.

• The OECD’s “Student Agency for 2030” places agency and co-agency at the centre of future-ready education.

• Duckworth’s grit research ties long-term achievement to self-driven perseverance toward one’s own goals.

• The collapse of fixed career paths (Article 2) removes the external structure that once substituted for internal agency — making it decisive.

• Self-regulated-learning research (Zimmerman; Panadero) shows the self-directing disposition is teachable and drives independent achievement.

How to build with it:

• Give students real ownership: genuine choices, genuine responsibility, and the standing expectation that they initiate.

• Flip the hidden curriculum from compliance to initiative — reward the student who starts, not only the one who obeys.

• Use self-directed projects with real stakes so agency is practised, not merely preached.

• Teach self-authorship and internal locus of control as explicit, nameable capacities.

• Treat passivity as the real failure mode of an education — the extra who waits, not the protagonist who acts.

10. The Body as Base

Metaphor: Every genius you will ever be is running on a single piece of hardware you can either maintain or ruin — and there is no upgrade and no replacement.

Definition: This is the deliberate maintenance of the body that runs the mind.
It is sleep, nutrition, movement and the regulation of the nervous system.
It is the recognition that attention, memory and judgement are all downstream of biology.
It replaces the afterthought gym-slot with health as cognitive infrastructure.
It is the one machine no one else — and no AI — can maintain for you.
Neglect it and every higher faculty degrades; train it and they compound.

Why it holds:

• The exercise-and-cognition research (the BDNF/neurogenesis review) shows physical activity physically builds the brain’s capacity to learn and remember.

• Systematic reviews and cohort studies (ERIC; PLOS ONE) link physical activity to measurable gains in academic performance.

• Sleep research links sleep quality directly to wellbeing and academic outcomes.

• The attention literature shows focus is a fatigable biological resource, tied to the state of the body and nervous system.

• Embodied-cognition work confirms the mind is grounded in the body, not separable from it.

How to build with it:

• Teach health as a serious subject — the science and practice of sleep, nutrition, movement, stress regulation — not an afterthought.

• Build the school day to protect the substrate: sleep-aware timetables, real movement, regulated environments.

• Give students ownership of their own vitality as a lifelong performance advantage.

• Integrate movement and regulation into learning, not against it.

• Frame the body as the base on which every other faculty in this curriculum stands.

11. Capital & Power

Metaphor: In a storm it matters enormously whether you own the boat or are bailing water in someone else’s — and no one ever taught the bailers the difference.

Definition: This is fluency in money, ownership and the mechanics of wealth.
It is the difference between earning wages and owning assets that earn for you.
It is capital allocation, equity, compounding, risk, and financial independence.
It replaces the near-total absence of money from the curriculum — the silence that serves the already-wealthy.
It is the defence against an economy where returns concentrate in owners.
It is sovereignty in its most literal, material form.

Why it holds:

• Financial-literacy research (Kaiser and Lusardi; Hastings et al.) links financial capability causally to wealth, saving and real economic outcomes.

• OECD PISA data show large, unequal gaps in young people’s financial literacy — a measurable, addressable deficit.

• The macro literature shows AI shifting returns from labour toward capital, widening the owner-earner divide (Article 2, force 10).

• The winner-take-most dynamics of AI economies concentrate gains among owners of systems, raising the stakes of financial understanding.

• The GFLEC and World Bank evidence shows financial capability can be measured and built, especially in the young.

How to build with it:

• Teach money, ownership and capital universally and early — the mechanics of wealth, taught to everyone, not absorbed by the lucky few.

• Teach the crucial distinction between earning and owning, and the pathways from one to the other.

• Connect it to entrepreneurship (subject 5): building and owning value, not only selling labour.

• Use real, concrete practice — budgets, investments, equity, compounding — grounded in students’ actual futures.

• Treat financial illiteracy as an equity emergency and its remedy as a form of empowerment.

12. Why Anything At All

Metaphor: When the machines can do everything, the only question left standing is the one school never dared to ask: what is worth doing, and who do you want to be?

Definition: This is the deliberate cultivation of meaning, values and the examined life.
It is the capacity to answer, for oneself, why any of it matters.
It is ethics, purpose, character — the formation of a person, not just a worker.
It replaces the curriculum’s silence on meaning with its most important subject.
When work becomes optional and answers become free, meaning is the scarce resource.
It is the antidote to the nihilism and hedonic drift that abundance invites.

Why it holds:

• Damon’s research shows a developed sense of purpose is a powerful driver of engagement, resilience and wellbeing in the young.

• Seligman and Adler’s positive-education work argues meaning and wellbeing can and should be cultivated deliberately.

• The Frontiers adolescent research links purpose in life to protection against depression and to flourishing.

• Character-education frameworks (the Jubilee Centre; Oxford) show virtue and purpose can be taught as serious, structured disciplines.

• Self-determination theory ties durable motivation to purpose and relatedness — internal sources that survive the automation of work.

How to build with it:

• Put meaning, ethics and the examined life at the centre of the curriculum, taught seriously and honestly.

• Help each student build an internal source of meaning independent of any job or external validation.

• Teach the great questions — what is a good life, what do I owe others, what is worth wanting — as core, not enrichment.

• Connect learning to contribution and purpose, so effort is anchored in something that matters.

• Treat the capacity to generate one’s own meaning as a survival skill for an age of abundance.

Cluster IV — The Arena

13. The Power to Move People

Metaphor: When every worker has a genius in their pocket, the one irreplaceable act is making a room of humans believe you, follow you, and act.

Definition: This is the mastery of human influence — rhetoric, story, persuasion, negotiation.
It is the ability to move people with words, spoken and written.
It is the oldest power in the world and the one machines do not hold.
It replaces the stale, decontextualised “English composition” with the live art of influence.
It is dominance in its legitimate form: the capacity to make your case and win it.
Human-to-human persuasion does not get automated, and it decides almost everything.

Why it holds:

• The persuasion literature (Petty and Cacioppo’s Elaboration Likelihood Model) establishes influence as a structured, learnable discipline, not a mysterious gift.

• Deming’s research shows the labour-market premium on social and communication skills rising as cognition automates.

• The “Writing Next” meta-analysis (Graham and Perin) shows effective writing is teachable through identifiable, evidence-based strategies.

• ETS’s work frames communication as a core future-ready skill for education and work.

• The complementarity argument implies human-to-human influence is exactly the layer machines cannot occupy.

How to build with it:

• Teach rhetoric, story and persuasion rigorously — writing that lands, speaking that moves, arguing that wins.

• Give students repeated experience persuading real audiences and negotiating real stakes.

• Teach the structure of influence — the routes to persuasion, the anatomy of a compelling story — explicitly.

• Assess the ability to move an audience, not just to produce grammatically correct text.

• Frame it as legitimate power: the capacity to make your ideas prevail in a world of competing voices.

14. Command of People

Metaphor: The future economy is not a solo sport played against machines; it is a team sport, and the captains — the people others trust — take the field.

Definition: This is the capacity to build trust, teams and alliances.
It is leadership, emotional regulation, reading a room, and drawing the best from others.
It is the relational intelligence that turns a group of people into a force.
It replaces the near-absence of leadership and collaboration from formal schooling.
As cognition automates, the relationship-dense work rises to the top of the value chain.
The most automation-proof thing about a person is their effect on other people.

Why it holds:

• Deming’s landmark research shows jobs combining cognitive and social skills grew most, and that social-skill returns are rising.

• Social-emotional-learning research (CASEL’s meta-analysis; the Learning Policy Institute’s twelve meta-analyses) shows these capacities are teachable and predict long-run success.

• Cooperative-learning research demonstrates collaboration skills can be built and that they raise achievement.

• The WEF consistently ranks leadership, influence and teamwork among core rising competencies.

• The complementarity literature places the relational layer squarely in the human, non-automatable domain.

How to build with it:

• Teach leadership, teamwork and emotional regulation explicitly, through real responsibility for real groups.

• Build social-emotional skills as economic preparation, not just pastoral care.

• Give students repeated experience leading, collaborating and building trust — assessed like any core skill.

• Teach the reading of people and situations — the tacit relational intelligence — through practice and feedback.

• Protect and expand the deeply human, relational core of school; it is now the competitive edge.

15. Playing for Real

Metaphor: You cannot learn to swim in a classroom about swimming; at some point someone has to throw you in water that is genuinely deep.

Definition: This is high-agency learning under genuine stakes.
It is learning by doing real things with real consequences, not simulating them for a grade.
It is competition, risk, ownership of outcomes, and the intensity that only reality supplies.
It replaces the school-as-waiting-room — years of rehearsal for a life that starts “later.”
It treats the student as a player in the world now, not a spectator preparing to enter it.
Stakes are the strongest teacher there is, and school has anaesthetised itself against them.

Why it holds:

• Self-determination theory shows autonomy and authentic challenge drive the deepest engagement and motivation.

• Project-based and deeper-learning evidence (MDRC’s PBL review; AIR) shows real, consequential projects produce stronger outcomes than abstract exercises.

• Apprenticeship and work-based-learning research (OECD; ILO; Cedefop) shows learning under real conditions builds durable, transferable competence.

• The exemplar models (High Tech High; the microschool and Montessori evidence) sustain motivation precisely through real projects and genuine agency.

• The entrepreneurship literature shows building real things with real feedback develops capability faster than description.

How to build with it:

• Build the curriculum around real projects with real stakes — actual audiences, actual consequences, actual ownership.

• Use competition, risk and genuine challenge as motivating forces, not things to shield students from.

• Replace rehearsal-for-later with participation-now: students do real work in the real world as they learn.

• Draw on apprenticeship and project-based models where learning and doing are the same act.

• Treat the ability to perform under real stakes as the point of school, not a distant reward for finishing it.

16. Who Holds the Reins

Metaphor: However good the autopilot, someone with a name, a conscience and their hands near the controls has to be the captain — and knowing when to grab the controls is the whole job.

Definition: This is the mastery of the human-machine contract.
It is the judgement of when to delegate to a machine and when a human must decide.
It is knowing where accountability must stay human, and taking it.
It replaces the naive extremes — total techno-surrender and total refusal — with governed partnership.
It is the ethics of autonomy: responsibility for what the systems you command actually do.
As machines do more, this distinctly human role of holding the reins grows more important, not less.

Why it holds:

• The WEF’s AI-agents governance work stresses accountability, oversight and keeping humans answerable for agent behaviour.

• Hadfield and Koh’s work shows agents require human institutions and accountability structures to function safely.

• The US Department of Education’s guidance insists on humans staying “in the loop” and accountable for decisions affecting students.

• Human-AI interaction research (Amershi et al.; Bansal et al.) shows appropriate reliance — knowing when to trust the machine — is a hard, learnable skill.

• The verification and hallucination evidence makes human judgement the last line of defence against confident machine error.

How to build with it:

• Teach the ethics and practice of delegation — when to trust, when to override, who is accountable for the outcome.

• Build the judgement to know the machine’s limits and to own the results of its actions.

• Make moral reasoning about automated systems a core, examined part of the curriculum.

• Practise accountable decision-making with AI in the loop, where the human owns the outcome.

• Frame responsibility as a distinctly human role that grows, not shrinks, as machines do more — the reins stay in human hands.

So what: a whole person, not a filled vessel

Lay the sixteen out together and you do not see a syllabus. You see a person. A mind that can concentrate, discern truth, see whole systems and imagine futures. A builder who can create value, command machines, exercise taste and shape matter. A self that owns its agency, its body, its capital and its meaning. And a human who can move people, lead them, act under real stakes, and hold the reins of the systems they command. That is not a list of things to know. It is a description of a sovereign — someone who owns their own reality rather than being owned by the systems around them. The old curriculum filled a vessel. This one builds a person.

The objection will come immediately: this is idealistic, unassessable, impossible to timetable. It is none of those things. Every one of the sixteen rests on a research literature — cited above, drawn from a library of 179 primary studies — showing it is a real capacity that predicts real outcomes and can be deliberately built. Attention training works. Financial education works. Metacognition, social-emotional learning, project-based learning, apprenticeship, character education, entrepreneurship — all have evidence bases showing schools can develop them. The barrier has never been feasibility. The barrier is that we are still pouring our resources into teaching the horse-shoeing, out of habit, and calling it rigour.

None of this discards knowledge. You cannot discern truth, see systems, exercise taste or hold the reins from an empty head — the cognitive science is emphatic that every one of these faculties is built on deep, rich, well-organised knowledge. What changes is the purpose of the knowledge. It stops being the thing you are graded on reproducing and becomes the material from which the faculties are forged. History becomes the laboratory of Building Futures. Literature becomes the training ground of Taste and the Power to Move People. Science becomes Systems Sight and Mastery of Matter. The subjects don’t vanish; they are conscripted into the service of building a person instead of filling one.

So burn the timetable — not the knowledge, the timetable, the century-old assumption that education is the transmission of content to be tested and forgotten. Replace it with the deliberate construction of the sixteen faculties that a human needs to be sovereign in a world of abundant intelligence. The machines are going to be extraordinary. The only question that matters is whether the humans standing next to them are extraordinary too — or whether we spent their childhoods preparing them, with great care and enormous expense, for the one race they were always going to lose. Build owners. The rest is horse-shoeing.

Built on a purpose-built ENSI research library of 179 primary documents across 30 research angles — from the neuroscience of attention and the science of learning to the economics of the agentic labour market — spanning the OECD, UNESCO, the WEF, NBER, Stanford HAI, the EEF, CASEL, the Jubilee Centre and dozens more. Part three of the ENSI “Education for the Agentic Age” series.

Consider what has already happened to the ground under our feet. The World Economic Forum’s employers, surveyed across the whole world economy, expect 170 million new roles to be created and 92 million to be destroyed by 2030 — a churn of a quarter of all jobs inside a single presidential term. Researchers at OpenAI estimate that around eight in ten workers have at least ten percent of their tasks exposed to large language models. Whole categories of white-collar work — the “good jobs” a generation of parents told their children to aim for — are being unbundled into tasks, and the routine tasks are being handed to machines that cost cents and never sleep. The most-hired job titles of 2035 are, quite literally, not in today’s dictionary.

And yet in most classrooms on most mornings, the single most rewarded activity is a child silently reproducing information that a free chatbot can generate perfectly in four seconds. We are drilling children, at enormous expense, in precisely the capabilities we have just finished automating. It is as if, on the eve of the automobile, we had doubled down on teaching everyone to shoe horses — and graded them on the neatness of the nails.

This is not an argument that school is obsolete. It is the opposite. In a world where machines can do so much, what only humans can do becomes the most valuable thing on Earth — and school is the institution we built to develop exactly that. The tragedy is not that we still have schools. It is that our schools are optimising for the wrong human. They are still producing the compliant executor when the economy is screaming for the sovereign creator, the one who directs the machines rather than competing with them.

So this article does something specific. It sets aside, for now, the question of how to run AI in the classroom — that was the previous piece. It asks the prior question: what is the world actually going to demand of a person, and what does that demand imply about what school must build? It traces fourteen forces reshaping work and society as intelligence becomes abundant, and it reads each one for its educational instruction. Because the fourteen forces do not point in fourteen directions. They converge, with startling consistency, on a small set of durable human faculties — the capacities that get more valuable as machine intelligence gets cheaper.

The logic is worth stating plainly before we begin, because everything follows from it. When any capability becomes abundant and cheap, its economic value falls, and the value migrates to whatever is now the binding constraint. For two centuries the scarce, valuable thing was the ability to process information and execute procedures — so that is what school built and the economy paid for. Machines have now made that abundant. The new scarce things are the ones the machines still cannot supply: the judgement to decide what is worth doing, the taste to tell good from merely plausible, the agency to start without being told, the trust to lead other humans, the attention to think deeply, and the character to know why any of it matters. These do not appear on a report card. They are about to become the whole game.

Read together, the forces make a case that is almost embarrassingly clear: the curriculum should be reorganised around human faculties that appreciate, not bodies of content that depreciate. Geography as a list of capitals, history as a list of dates, “computer studies” as a set of menu commands — these are not wrong to know, but they are catastrophically wrong as the organising principle of an education, because they optimise for the exact cognition we have automated. The organising principle has to become the person we are trying to build.

At the end, those faculties are named — the sixteen subjects of a curriculum for owners, which the third article in this series defines in full. But the names only earn their place if the forces demand them. So here are the forces, and here is what each one is quietly instructing us to teach.

The main points, in short

1. The Great Inversion — execution collapses in value; judgement, taste and direction become the scarce, priced things.

2. The Task Is the Unit, Not the Job — automation dissolves jobs into tasks and recombines them; adaptability beats any fixed role.

3. From Employee to Orchestrator — the org unbundles into humans directing fleets of agents; commanding intelligence is the new literacy.

4. The Collapsing Half-Life of Skills — what you know expires faster than ever; learning-to-learn becomes the master competence.

5. The Social-Skills Premium — as cognition automates, the returns to trust, persuasion and teamwork rise.

6. The Return of Agency — with no fixed path, the self-starter wins and the instruction-follower stalls.

7. Taste as the New Scarcity — when content is infinite and free, the ability to judge quality is the moat.

8. Verification Becomes a Job — in a flood of confident synthetic content, discernment is a core economic function.

9. The Entrepreneurial Default — more people must create value directly rather than fill a pre-made slot.

10. The Winner-Take-Most Economy and the Ownership Divide — returns concentrate in owners; financial and capital literacy becomes survival.

11. The Meaning Crisis — as work stops being the source of identity, purpose becomes the scarce psychological resource.

12. The Body as Competitive Advantage — attention, energy and health become the substrate everything else runs on.

13. The Human-in-the-Loop Mandate — accountability for what machines do becomes a distinct, non-delegable human role.

14. The Sovereignty Imperative — the master-shift is from being managed by systems to owning your relationship with them.

The fourteen forces — and what each one tells us to teach

1. The Great Inversion

Metaphor: For centuries the ladder to the top was made of rungs called “doing the work”; the machines just took the rungs, and left the summit.

Definition: The industrial economy paid for the processing and execution of information.
Reading, calculating, drafting, filing, coding to spec — these were scarce, so they were valuable.
Machine intelligence has made all of them abundant, and abundance destroys price.
The value does not vanish; it migrates up, to whatever is now scarce.
What is scarce now is deciding what to do, judging whether it is good, and taking responsibility for it.
Execution is being inverted from the prize into the commodity, and judgement is taking its place.

Why it holds:

• Autor’s foundational work shows automation substitutes for routine tasks but raises the value of the complementary human ones — judgement, flexibility, problem-solving.

• Brynjolfsson’s “Turing Trap” argues the economic prize lies in AI that augments distinctly human capabilities rather than imitating and replacing them.

• The OpenAI exposure study finds it is precisely the routine, procedural tasks — the old core of “good jobs” — that are most exposed to LLMs.

• Acemoglu’s macroeconomics of AI locates gains in reallocation toward tasks where humans remain complementary, not in doing the old tasks faster.

• Deming and others document rising wage returns to non-routine, human-complementary skills as routine cognitive work is automated.

How to build with it:

• Reorient the curriculum from execution to judgement: teach students to decide what is worth doing and to evaluate quality, not just to produce.

• Stop rewarding flawless reproduction of automatable procedures as the summit of achievement.

• Build the faculties that sit above execution — synthesis, taste, discernment, direction (Article 3’s Sovereign Mind and Builder clusters).

• Treat “the machine can do this now” as a signal to move the learning goal up a level, not to lower the bar.

• Make the human’s comparative advantage — judgement under uncertainty — the explicit thing school develops.

2. The Task Is the Unit, Not the Job

Metaphor: Automation doesn’t swallow whole professions in one bite; it nibbles them apart task by task, then recombines the leftovers into jobs that never existed.

Definition: “Will AI take my job?” is the wrong question because a job is a bundle of tasks.
Automation hits the tasks, not the title, and it hits them unevenly.
Some tasks in every job vanish; others become more valuable because the machine amplifies them.
The jobs of the next decades are recombinations — human tasks re-bundled around what machines can’t do.
This makes any specific, fixed occupational training a depreciating asset.
What appreciates is the adaptability to keep recombining as the boundary moves.

Why it holds:

• The OpenAI/Eloundou study is explicitly task-based: exposure varies task by task within every occupation, not job by job.

• The ILO’s analysis finds generative AI mostly augments — transforming task mixes — rather than wholesale automating occupations.

• The WEF’s data show simultaneous large-scale creation and destruction, i.e. recombination, not simple net loss.

• Autor’s “rebuild middle-class jobs” argument is precisely about recombining tasks so human expertise is extended, not erased.

• Historical evidence (Autor’s “why are there still so many jobs”) shows automation repeatedly destroyed tasks while creating new occupations around the remaining human ones.

How to build with it:

• Teach adaptability and transfer as first-class goals — the ability to move into new task-bundles, not mastery of one fixed role.

• Build broad, deep, transferable foundations (knowledge-rich, per the cognitive science) plus the meta-skill of re-learning fast (force 4).

• Orient career preparation around durable human tasks (judgement, creativity, relationship, orchestration), not job titles that may not survive.

• Use project work that forces students to recombine skills for novel problems, mirroring the real dynamic.

• Kill the assumption that education ends with a credential for a job; build for a life of continuous recombination.

3. From Employee to Orchestrator

Metaphor: The most valuable person in the room is shifting from the one who plays an instrument beautifully to the one who conducts the whole orchestra of machines.

Definition: The firm is being unbundled by agents that can plan, act and transact.
Increasingly, one human plus a fleet of AI agents does what a team of humans used to do.
The human’s role in that arrangement is not to be a better executor than the agents.
It is to direct them: to specify, delegate, verify, and hold the whole system to a standard.
This is a genuinely new literacy — the ability to command intelligence, not just to possess it.
The person who can orchestrate is worth many people who can only execute.

Why it holds:

• The “agentic economy” research (Microsoft Research; the ten-principles framework) describes an economy where assistant and service agents transact on behalf of humans and firms.

• Hadfield and Koh’s NBER work on “an economy of AI agents” analyses agents deploying across markets and the institutions they require to be governed.

• Stanford HAI’s AI Index documents the surge of investment and adoption pushing agentic capability into real workflows.

• Brynjolfsson, Li and Raymond’s field study shows the biggest productivity gains flow to workers who effectively direct AI in their tasks.

• Prompt-pattern and human-AI-interaction research (White et al.; Amershi et al.) shows orchestration is an engineerable, learnable skill, not an innate one.

How to build with it:

• Make orchestration a taught skill: specification, delegation, verification, and taste in judging machine output (Article 3’s “Commanding the Swarm”).

• Move students from “produce the answer yourself” to “direct and verify a system that produces it, and own the result.”

• Teach the judgement to know when to delegate and when to do it yourself (linked to the human-in-the-loop mandate, force 13).

• Practise real human-AI teaming on substantial projects, with the human accountable for quality.

• Treat AI fluency as the new baseline literacy — as fundamental as reading — not an optional extra.

4. The Collapsing Half-Life of Skills

Metaphor: You used to buy your knowledge like a house, once, for life; now you rent it, and the lease keeps getting shorter.

Definition: The specific, technical content of expertise now expires at accelerating speed.
A tool learned today may be obsolete before a four-year degree finishes.
This does not make knowledge worthless — deep foundations matter more than ever (see force 5 and the cognitive science).
But it makes any fixed body of applied content a wasting asset.
The one competence that never depreciates is the ability to learn — fast, deliberately, alone.
Learning-to-learn stops being a nice-to-have and becomes the master skill of a working life.

Why it holds:

• The WEF’s Future of Jobs data show employers ranking analytical thinking, resilience, curiosity and lifelong learning among the fastest-rising skills.

• The OECD Learning Compass 2030 is built around exactly this: transformative competencies and the capacity to keep learning, not a fixed knowledge list.

• Metacognition and self-regulated learning are, per the EEF and Hattie, among the highest-leverage and most transferable of all educational outcomes.

• Zimmerman’s and Panadero’s models show self-regulated learning is teachable and drives the ability to acquire new skills independently.

• The task-recombination dynamic (force 2) guarantees repeated re-skilling across a career, making the meta-skill economically essential.

How to build with it:

• Put metacognition and learning-to-learn at the centre of the curriculum — planning, monitoring, self-testing, reflection — taught explicitly (Article 3’s Sovereign Mind).

• Teach the science of learning to the learners themselves so they can run their own development for life.

• Build the disposition of self-directed learning through genuine autonomy and mastery experiences, not just compliance.

• Assess the ability to learn something new under time pressure, not only the possession of pre-taught content.

• Frame school’s deliverable as “a person who can teach themselves anything,” not “a person who has been taught these things.”

5. The Social-Skills Premium

Metaphor: When every worker has a genius in their pocket, the rare and priceless thing becomes the ability to make other humans trust you, follow you, and work with you.

Definition: As machines absorb cognitive tasks, the human-to-human layer rises in value.
Persuading, negotiating, leading, collaborating, reading a room — machines do not do these.
They are the connective tissue of every organisation and every market.
And the data show their wage premium climbing precisely as routine cognition automates.
Relationship is becoming not a soft skill but a hard economic advantage.
The most automation-proof thing about a person is their effect on other people.

Why it holds:

• Deming’s landmark research shows the labour market increasingly rewards social skills, and that jobs combining cognitive and social skills grew most.

• The WEF consistently ranks people-skills — leadership, influence, collaboration — among the core rising competencies.

• The complementarity literature (Autor) implies the human relational layer is exactly what machines cannot substitute for.

• Social-emotional learning research (CASEL; the Learning Policy Institute’s twelve meta-analyses) shows these skills are teachable and predict long-run outcomes.

• The persuasion literature (Petty and Cacioppo) establishes influence as a structured, learnable discipline, not a mysterious gift.

How to build with it:

• Teach rhetoric, persuasion, leadership and collaboration explicitly and rigorously (Article 3’s “Arena” cluster).

• Make real teamwork, negotiation and human influence central assessed activities, not extracurricular garnish.

• Invest in social-emotional learning as economic preparation, not just pastoral care.

• Give students repeated experience leading, persuading and building trust in front of real audiences.

• Protect and expand the deeply human, relational side of school — it is now the competitive edge.

6. The Return of Agency

Metaphor: When the conveyor belt stops, the people who only knew how to stand on it are stranded; the ones who can walk go anywhere.

Definition: The industrial economy ran on a predetermined path: do these steps, get this life.
School was built to produce people who could follow that path reliably.
The path is dissolving — careers are self-authored, non-linear, improvised.
In that world the decisive trait is agency: the disposition to initiate without being told.
The instruction-follower waits for a conveyor belt that is no longer coming.
The self-starter, the one who acts as the owner of their situation, thrives.

Why it holds:

• The OECD’s “Student Agency for 2030” places agency and self-direction at the centre of future-ready education for exactly this reason.

• Self-determination theory (Ryan and Deci) shows autonomy is a root of motivation, persistence and wellbeing — the fuel of self-authored lives.

• Duckworth’s grit research links long-term achievement to self-driven perseverance toward one’s own goals.

• The collapse of fixed career paths (forces 2 and 4) removes the external structure that used to substitute for internal agency.

• Entrepreneurship research (force 9) shows initiative and opportunity-recognition as learnable dispositions that increasingly determine outcomes.

How to build with it:

• Build agency and ownership deliberately: give students real choices, real responsibility, and the expectation that they initiate (Article 3’s “Living as an Owner”).

• Shift the hidden curriculum from compliance to initiative — reward the student who starts, not only the one who obeys.

• Use self-directed projects with genuine stakes so agency is practised, not just praised.

• Teach an internal locus of control and self-authorship as explicit, nameable capacities.

• Treat passivity, not disruption, as the real failure mode of an education for this economy.

7. Taste as the New Scarcity

Metaphor: When anyone can generate a thousand paintings a minute, the priceless person is the one who can point at the one worth keeping.

Definition: Generative machines make production of content essentially free and infinite.
When supply of anything explodes, the bottleneck moves to selection.
The scarce, valuable act becomes judging which of the infinite options is actually good.
That judgement — taste — is hard-won, tacit, and deeply human.
It cannot be prompted into existence; it is built through exposure, practice and discernment.
In a world drowning in plausible mediocrity, taste is the moat.

Why it holds:

• Generative models are engines of the probable — they regress to the average of their training data, so distinguishing the excellent from the merely plausible is a human residual.

• The WEF ranks creative thinking and analytical judgement among the fastest-rising skills as content generation is automated.

• The complementarity argument (Autor; Brynjolfsson) implies value concentrates in the human acts machines can’t do — curation and judgement chief among them.

• Research on AI dependence and creativity (the Frontiers study) warns that offloading judgement to the machine erodes the very taste that is now scarce.

• Expertise research (Ericsson) shows refined judgement is built through deliberate practice — it is trainable, but only through effortful exposure.

How to build with it:

• Teach taste and origination as explicit disciplines: the judgement to recognise quality and the courage to make something genuinely new (Article 3’s “Taste & Origination”).

• Immerse students in the excellent across domains, so their internal standard rises.

• Assess judgement — “which of these is best, and why” — not just production.

• Use AI to generate options and make students the editors, exercising and building discernment.

• Reward the rare, the surprising and the well-judged over the fluent and the average.

8. Verification Becomes a Job

Metaphor: In a city where every stranger might be a brilliant con artist, the most valuable citizen is the one who can tell truth from performance.

Definition: Machine intelligence has made confident, fluent, plausible content nearly free.
Some of it is true and some is invented, and the two are indistinguishable on the surface.
This turns verification from a background chore into a foreground economic function.
Every field will need people who can tell what is real, sourced and load-bearing.
The skill is epistemic: sourcing, calibration, probabilistic reasoning, disciplined doubt.
In an ocean of synthetic plausibility, the verifier is indispensable.

Why it holds:

• The detector studies (Weber-Wulff; the Stanford bias study) show that surface fluency carries no signal of truth — verification cannot be automated away.

• Microsoft Research finds that trust in AI displaces scrutiny, so deliberate verification skills become a scarce corrective.

• The dark-patterns and attention-economy literature shows the information environment is actively engineered to mislead, raising the premium on discernment.

• The Elaboration Likelihood Model explains why plausible fluency persuades under low effort — making trained skepticism economically valuable.

• The persistence of hallucination as a core failure mode of LLMs guarantees ongoing human demand for verification.

How to build with it:

• Teach discernment and verification as a core subject: sourcing, evidence, calibration, probabilistic thinking, bullshit-detection (Article 3’s “Discernment”).

• Make “check before you trust” a trained reflex applied to human and machine claims alike.

• Assess the ability to evaluate evidence and detect manipulation, not just to recall or produce.

• Ground verification in deep domain knowledge — you can only fact-check what you understand.

• Frame epistemic hygiene as both a civic duty and a marketable skill.

9. The Entrepreneurial Default

Metaphor: When there are fewer pre-built chairs to sit in, more people have to learn to build their own.

Definition: The stable, salaried slot inside a large organisation is becoming less central.
Value increasingly accrues to those who can create it directly — spot a need, build a solution, capture the return.
AI radically lowers the cost of building, so the leverage of a single creator explodes.
This pushes entrepreneurial capability from a niche trait toward a general necessity.
Not everyone will start a company, but everyone will have to think like a value-creator.
The default posture shifts from “find a job” to “make something people want.”

Why it holds:

• The OECD’s entrepreneurship-education work argues the entrepreneurial mindset — opportunity recognition, initiative, value creation — is teachable and increasingly essential.

• The agentic-economy research shows AI collapsing the cost and headcount needed to build and ship, amplifying individual leverage.

• Amit and Zott’s value-creation framework formalises how novel resource configurations — the entrepreneur’s core act — generate and capture value.

• Brynjolfsson, Li and Raymond, and Noy and Zhang, show AI most amplifies those who direct it to create, compressing the advantage of incumbents.

• NBER work on business dynamism underscores how much economic vitality depends on new value creation — and how costly its decline is.

How to build with it:

• Teach value creation and entrepreneurship as a core subject, from the builder’s chair, not as an elective (Article 3’s “The Value Engine”).

• Have students actually make and ship things people want — real projects, real audiences, real feedback.

• Teach the mechanics of how value is created and captured, not just idea-generation.

• Cultivate the entrepreneurial dispositions — initiative, resilience, opportunity-spotting — as trainable habits.

• Reward building over describing; make “you made something real” the highest form of schoolwork.

10. The Winner-Take-Most Economy and the Ownership Divide

Metaphor: In a storm, it matters enormously whether you own the boat or are bailing water for someone who does.

Definition: Digital and AI-driven economies concentrate returns dramatically.
Leverage flows to those who own — equity, assets, systems — not those who merely earn wages.
As AI compresses the value of labour, the gap between owners and earners widens.
A person who understands capital, ownership and allocation can ride the wave.
A person financially illiterate is exposed to it, with no defence and no upside.
The single most consequential divide of the coming decades may be ownership itself.

Why it holds:

• The macro literature (Acemoglu; the labour-share debate) shows AI shifting returns from labour toward capital and owners.

• Financial-literacy research (Kaiser and Lusardi; Hastings et al.) links financial capability causally to wealth, saving and real economic outcomes — and documents how little of it is taught.

• OECD PISA data show large, unequal gaps in young people’s financial literacy across and within countries.

• The winner-take-most dynamics of platform and AI economies concentrate gains among owners of the systems.

• The near-total absence of money, capital and ownership from standard curricula leaves this divide entirely unaddressed.

How to build with it:

• Teach money, capital and ownership as a core subject — the mechanics of wealth, equity versus wages, allocation, sovereignty (Article 3’s “Capital & Power”).

• Make financial and economic literacy universal, starting early, as a basic defence against the ownership divide.

• Teach the distinction between earning and owning, and the pathways from one to the other.

• Connect it to entrepreneurship (force 9): building and owning value, not just selling labour.

• Treat financial illiteracy as the equity emergency it is — the gap that compounds across a lifetime.

11. The Meaning Crisis

Metaphor: For a century work told people who they were; when the machine can do the work, the mirror goes blank.

Definition: In the industrial and knowledge economies, work was the main source of identity and structure.
“What do you do?” was a question about who you are.
As machines absorb more of the doing, that source of meaning is destabilised.
For some, work becomes optional; for many, it becomes precarious or fluid.
Either way, the psychological load shifts onto a scarcer resource: an internal sense of purpose.
A generation that cannot manufacture its own meaning is exposed to nihilism and despair.

Why it holds:

• Damon’s research shows a developed sense of purpose is a powerful protective and motivating force in young people.

• Seligman and Adler’s positive-education work argues wellbeing and meaning must be cultivated deliberately, as explicit aims.

• The Frontiers adolescent research links purpose in life to resilience against depression and to wellbeing.

• Self-determination theory ties durable motivation to purpose and relatedness — internal sources that survive the automation of work.

• Rising adolescent mental-health strain, set against the attention economy (CIGI), makes the deliberate cultivation of meaning urgent, not optional.

How to build with it:

• Put meaning, purpose and character at the centre of the curriculum, not the margins (Article 3’s “Why Anything At All”).

• Help students build an internal source of meaning independent of any specific job or external validation.

• Connect learning to contribution and purpose, so effort is anchored in something that matters to them.

• Teach the examined life — values, ethics, the question of what a good life is — as core, not enrichment.

• Treat the capacity to generate one’s own meaning as a survival skill for an age of abundance and disruption.

12. The Body as Competitive Advantage

Metaphor: You can own the finest racing car ever built, but if the engine is starved of fuel it loses to a bicycle.

Definition: Every cognitive faculty in this article runs on a biological substrate.
Attention, memory, judgement and creativity are all downstream of a brain and a nervous system.
That substrate is built and maintained by sleep, movement, nutrition and regulation.
In an economy that prizes cognition, the health of the substrate becomes a competitive edge.
And it is one of the few advantages a machine categorically cannot supply for you.
Neglect the body, and every higher faculty degrades; train it, and they compound.

Why it holds:

• The exercise-and-cognition research (the BDNF/neurogenesis review) shows physical activity physically builds the brain’s capacity for memory and learning.

• Systematic reviews and cohort studies (ERIC; PLOS ONE) link physical activity to measurable gains in academic performance.

• Sleep research links sleep quality to wellbeing and academic outcomes — a direct lever on cognitive capacity.

• The attention research (challenge 1 of Article 1) shows focus is a fatigable biological resource, tied to the state of the nervous system.

• Embodied-cognition work shows thinking itself is grounded in bodily action, not floating above it.

How to build with it:

• Teach embodied mastery and health as a serious subject — sleep, nutrition, movement, the nervous system — not an afterthought gym slot (Article 3’s “The Body as Base”).

• Treat physical vitality as cognitive infrastructure and build the school day around protecting it.

• Give students ownership of their own health as a lifelong performance advantage.

• Integrate movement and regulation into learning, not against it.

• Frame the body as the one machine each person must maintain themselves — no delegation possible.

13. The Human-in-the-Loop Mandate

Metaphor: However good the autopilot, someone with a name and a conscience still has to be legally, morally in the captain’s seat.

Definition: As agents act in the world, the question of who is responsible sharpens.
Machines can decide and execute, but they cannot be accountable — accountability is human.
Every consequential automated decision needs a human who owns the outcome.
This creates a distinct, non-delegable role: the human who holds the reins.
It demands judgement about when to trust the machine and when to override it.
Far from disappearing, human responsibility becomes more concentrated and more important.

Why it holds:

• The WEF’s AI-agents governance work stresses accountability, oversight and the need for humans to remain answerable for agent behaviour.

• Hadfield and Koh’s work highlights the institutions agents require — including human accountability structures — to function safely.

• The US Department of Education’s guidance insists on keeping humans “in the loop” and accountable for decisions affecting students.

• Human-AI interaction research (Amershi et al.; Bansal et al.) shows appropriate reliance — knowing when to trust the machine — is a skill, and a hard one.

• The verification imperative (force 8) makes human judgement the last line of defence against confident machine error.

How to build with it:

• Teach the ethics and practice of delegation — when to trust, when to override, who is responsible (Article 3’s “Who Holds the Reins”).

• Build the judgement to know the limits of the machine and to take ownership of outcomes.

• Make moral reasoning about automated systems a core, examined part of the curriculum.

• Practise accountable decision-making with AI in the loop, where the human owns the result.

• Frame responsibility as a distinctly human role that grows, not shrinks, as machines do more.

14. The Sovereignty Imperative

Metaphor: In every era there are those who are programmed by the system and those who program it; the gap between them has never been wider.

Definition: Underneath all thirteen forces sits a single master-shift.
The systems around us — algorithmic, economic, informational — are becoming vastly more powerful.
A person can relate to them in one of two ways: as a subject, or as a sovereign.
The subject is managed by the feed, the platform, the employer, the algorithm.
The sovereign owns their attention, their judgement, their capital, their choices.
Education’s deepest task is to produce sovereigns — owners of their own reality — not subjects.

Why it holds:

• The attention-economy literature (James Williams; dark patterns; the OECD) shows systems engineered to capture and direct human behaviour against the person’s interest.

• The ownership-divide evidence (force 10) shows the widening gap between those who own systems and those owned by them.

• The agency and self-determination research (forces 6, 1) establishes ownership of one’s choices as a learnable, decisive disposition.

• The sovereignty framing unifies the appreciating faculties: attention, discernment, agency, capital, meaning are all forms of self-ownership.

• ENSI’s broader thesis — the agentic era rewards those who direct intelligence rather than serve it — makes sovereignty the organising aim.

How to build with it:

• Make sovereignty the organising purpose of the whole curriculum — every faculty is a form of owning one’s own mind, work and life.

• Teach students to see the systems acting on them and to take ownership of their relationship with each.

• Frame the sixteen faculties (below) as the components of a sovereign person.

• Reject an education that produces compliant subjects; aim explicitly at owners.

• Sell it to students as power: the century belongs to those who own their attention, judgement and capital.

So what: the sixteen faculties the forces demand

Read the fourteen forces as a single instruction and they say one thing. Reorganise education around the human faculties that appreciate as intelligence becomes abundant — and stop organising it around the content that depreciates. The forces do not ask for better geography lessons or an extra coding class. They ask for a different kind of human: focused, discerning, original, entrepreneurial, financially sovereign, relationally powerful, healthy, purposeful, and able to command machines rather than compete with them.

Named as subjects, the faculties the forces demand fall into four clusters. The Sovereign Mind — Cognitive Sovereignty (trained attention), Discernment (verification and truth), Systems Sight (synthesis), and Building Futures (foresight) — the inner faculties no machine can rent. The Builder — The Value Engine (entrepreneurship and value creation), Commanding the Swarm (orchestrating AI), Taste & Origination (judgement and originality), and Mastery of Matter (making real things) — the capacities that turn intelligence into value. The Sovereign Self — Living as an Owner (agency), The Body as Base (health), Capital & Power (money and ownership), and Why Anything At All (meaning and the examined life) — the operating system of a free person. And The Arena — The Power to Move People (rhetoric), Command of People (leadership), Playing for Real (learning under stakes), and Who Holds the Reins (the human-machine contract) — the capacities for acting among people and machines.

That is the curriculum for owners: sixteen faculties, not a list of subjects to be examined and forgotten. Each maps directly onto a force that is already reshaping the world your students will inherit. This article has made the case for why — why the forces demand exactly these, and why the old organising principle has expired. The next article does the harder and more radical thing: it takes each of the sixteen and defines it in full — what it is, why it holds, and how to actually build it in a human being. Because naming the faculties is easy. Building them, deliberately, in a school designed for a world that no longer exists, is the real work — and it is the work that will decide whether the next generation inherits the machines, or is inherited by them.

Built on a purpose-built ENSI research library of 179 primary documents across 30 research angles — spanning the World Economic Forum, the OECD, the ILO, NBER, the World Bank, Stanford HAI, Microsoft Research and dozens more, from the economics of automation to the science of motivation and meaning. Part two of the ENSI “Education for the Agentic Age” series.

Every School Must Adopt AI — and Every School Must Defend the Mind Against It

In a lab at MIT, fifty-four students wrote essays while wearing EEG caps that read the electrical traffic of their brains. One group wrote unaided. One group could search the web. One group used ChatGPT. The machine-assisted essays were fluent, fast, competent — and the brains that produced them were the quietest in the room. The ChatGPT group showed the weakest neural connectivity of the three, the least coupling between the regions that bind memory to meaning. Then came the detail that should be printed on the wall of every ministry of education: minutes after finishing, most of the ChatGPT writers could not quote a single sentence from the essay they had just “written.” They had produced a document without forming a memory. The researchers gave the phenomenon a name — cognitive debt — and it is the debt an entire generation is about to take on without noticing.

Now hold that image next to a second one. In Edo State, Nigeria, secondary-school students met an AI tutor for six weeks of after-school sessions. The measured learning gain was equivalent to roughly two years of ordinary schooling — one of the most cost-effective education interventions the World Bank has ever recorded. At Harvard, students learning physics with a purpose-built AI tutor learned more than twice as much as peers in an acclaimed active-learning classroom, in less time. Same underlying technology. One version hollowed out the mind. The other supercharged it.

That is the whole problem in two pictures, and it dissolves the debate we keep having. The question is not whether to bring artificial intelligence into schools. That has already been decided — not by educators but by the economy students are graduating into. The question is how, because the distance between the MIT result and the Nigeria result is not a difference of technology. It is a difference of design. Artificial intelligence is simultaneously the most powerful tutor ever built and the most powerful cognitive off-switch ever built, and which one you get depends entirely on how you wire it into the day.

Adoption is not optional, and pretending otherwise is a form of malpractice. The World Economic Forum’s employers expect 170 million new roles to appear and 92 million to vanish by 2030. Researchers estimate that around eight in ten workers have at least some of their tasks exposed to large language models. A school that keeps AI out of the building in the name of protecting children is not protecting them; it is preparing them, with great care, for a world that will not exist. Fluency with these tools is becoming the difference between directing the machines and being managed by someone who does.

And yet adoption is genuinely dangerous, because the faculties AI erodes when it is misused — sustained attention, durable memory, the willingness to struggle, the habit of checking whether a claim is true — are precisely the faculties the new economy makes more valuable, not less. As routine cognition gets automated and nearly free, the premium shifts to the things machines cannot do: judgement, taste, originality, the capacity to hold a hard problem in your head long enough to crack it. If school hands those very capacities to the machine to “save time,” it is spending the endowment it exists to build.

So the reframe every educator, parent and policymaker needs is this. Stop asking whether students should be allowed to use AI. Start asking a sharper pair of questions: what must remain inside the student’s own head — non-negotiably, permanently — and how do we protect that while delegating everything else? Those two mandates run at once, in the same classroom, on the same day. Teach fluently with AI. Defend the mind from it. A school that does only the first produces confident incompetents who cannot function when the tool is taken away. A school that does only the second produces beautifully disciplined minds that are unemployable. The art is doing both.

This article is a field manual for doing both. It lays out twenty-four concrete challenges of bringing AI into education — the cognitive traps, the classroom-design problems, the systemic obstacles, and the case for urgency — and for each one it gives the evidence and the fix. It is grounded in a purpose-built library of roughly 180 primary studies, from the neuroscience of attention to the economics of the agentic labour market, and it is written for the person who has to make this real on Monday morning: the teacher, the head, the founder, the official who cannot afford either the techno-utopian brochure or the moral panic.

None of the twenty-four is unsolvable. But none solves itself, and several are actively made worse by the “obvious” response — banning the tools, or buying the tools and walking away. The pattern that runs through all of them is the same: AI should amplify a mind that has already done the work, never replace the work that builds the mind. Get the sequence right and you get Nigeria. Get it wrong and you get cognitive debt at national scale. Here is how to get it right, twenty-four times over.

The main points, in short

1. The Attention Collapse — AI arrives on the most distracting devices ever made; focus must be taught as a subject, not assumed.

2. The Memory Offload Trap — outsourcing memory to the machine weakens the very knowledge base that thinking runs on.

3. Cognitive Debt — using AI instead of struggling lowers brain engagement; struggle first, then augment.

4. Metacognitive Laziness — students hand their self-regulation to the chatbot; teach them to run themselves.

5. The Atrophy of Critical Thinking — the more we trust AI, the less we scrutinise it; build verification into every use.

6. The Illusion of Knowledge — access to answers feels like understanding; force closed-book explanation to break the spell.

7. Creative Dependence — reaching for the prompt first kills original thought; generate before you generate-with.

8. The Unguarded Chatbot — raw ChatGPT can lower exam scores; only guardrailed, Socratic tutors help.

9. The Tutor-Not-Answer-Key Problem — the whole game is designing AI that withholds the answer and coaches instead.

10. The Unclaimed Two-Sigma Prize — personalised tutoring finally works at scale, but only with the right deployment.

11. The Teacher’s New Job — the teacher shifts from deliverer of content to designer, coach and mentor.

12. The Death of the Take-Home Essay — unsupervised written homework is finished; move assessment to the process.

13. Integrity Without Surveillance — detectors fail and are biased; redesign tasks so cheating and not-learning become the same act.

14. The Equity Fork — AI will either widen the gap or close it; it helps novices most if we make it universal.

15. The Attention-Economy Adversary — school now competes with engineered addiction; teach the adversary and shrink its surface.

16. The Motivation Problem — when answers are free, why try; rebuild intrinsic motivation through autonomy, mastery and stakes.

17. The Meaning Problem — if a machine can do it, why learn it; reframe learning as self-formation, not task-completion.

18. The Knowledge-Is-Obsolete Fallacy — “just look it up” is a cognitive-science error; deep knowledge is what makes AI useful.

19. The Depth Tradeoff — AI makes shallow completion frictionless; reward depth, revision and defence of ideas.

20. The Developmental Mismatch — young brains are still building the machinery AI lets them skip; age-gate accordingly.

21. Data, Privacy and the Student Profile — children’s learning data is uniquely sensitive; govern it or don’t collect it.

22. The Hallucination Problem — confident falsehood is the default failure mode; teach distrust-and-verify as a reflex.

23. The Change-Management Wall — systems don’t adopt tools, people do; lead with teachers and evidence, not mandates.

24. The Cost of Not Adopting — refusing AI is also a decision, and it is the more expensive one.

The twenty-four challenges

1. The Attention Collapse

Metaphor: You cannot fill a bucket that has been drilled full of holes, and the smartphone is a drill.

Definition: Learning of any depth requires sustained, voluntary attention.
AI reaches students through the most attention-hostile devices ever engineered.
The same screen that runs the tutor runs the feed that fights the tutor.
Even a silent phone on the desk measurably drains working memory.
Divided attention does not slow learning; it prevents the encoding that makes learning stick.
So the first thing an AI school must protect is not data — it is focus.

Why it holds:

• Ward and colleagues found that the mere presence of one’s own smartphone, face-down and switched off, reduces available cognitive capacity — the mind spends effort not-attending to it.

• PISA 2022, analysed across nearly 80 education systems by the OECD, links in-class digital distraction to materially lower mathematics performance.

• The neuroscience of attention (Petersen and Posner; Corbetta and Shulman) shows focus is a limited, effortful resource run by specific, fatigable brain networks — not an infinite tap.

• Forster and Lavie show that when a task’s load is low, attention leaks outward to whatever is most salient — which, on a connected device, is engineered to be the interruption.

• UNESCO’s 2023 global monitoring report concluded that technology in classrooms is as likely to distract as to help unless its use is deliberately bounded.

How to build with it:

• Make attention a taught, assessed capability — deep-work blocks, single-tasking norms, and visible practice at holding focus, treated as seriously as literacy.

• Separate the surfaces: run AI tutoring on locked-down, single-purpose devices or modes, not on the same open browser that hosts the feed.

• Adopt phone-free defaults for the learning core of the day; PISA-grade evidence now supports it, and the burden of proof has flipped.

• Structure lessons around one hard thing at a time; design out the notification, the tab, the second screen.

• Begin sessions with a short attentional “warm-up” (a few minutes of focused breathing or silent reading) — cheap, and it primes the networks the lesson will tax.

2. The Memory Offload Trap

Metaphor: A crane that lifts every weight for you leaves you with arms that can no longer lift.

Definition: Human memory is not a filing cabinet you can empty into a device.
It is the substrate on which reasoning, comprehension and creativity actually run.
When we offload knowing-that to a machine, we stop building the internal schemas thinking needs.
The more we offload, the more we want to offload — it is a self-reinforcing habit.
Worse, retrieving from the machine feels like knowing, so we stop noticing the loss.
A mind with nothing in it has nothing to think with.

Why it holds:

• Sparrow, Liu and Wegner’s original “Google effect” experiments showed that when people expect information to remain available online, they remember it less well — they remember where to find it instead.

• Storm and colleagues found that using the internet to answer one question sharply increases the likelihood of reaching for the internet on the next — offloading is habit-forming.

• Ward’s work shows that searching online inflates people’s belief in their own internal knowledge, masking the erosion as it happens.

• The arXiv “memory paradox” analysis argues that even in an age of abundant AI, internalised knowledge remains necessary — expertise is compiled, not looked up.

• Cognitive-load research (Kirschner, Sweller, Clark) establishes that reasoning happens in working memory drawing on knowledge held in long-term memory; without the second, the first stalls.

How to build with it:

• Draw a hard line between knowledge you must internalise (the load-bearing facts, vocabulary and procedures of a domain) and knowledge you may offload — and defend the first fiercely.

• Front-load memory: students must be able to explain a concept from their own head before they are allowed to use AI to extend it.

• Use retrieval practice relentlessly — low-stakes quizzing, flashcards, teach-backs — as the antidote to the offload reflex (see challenge 6).

• Teach students the offload trap explicitly, so they can feel the difference between “I found it” and “I know it.”

• Reserve AI for the layer above mastered fundamentals — analysis, application, synthesis — not as a substitute for building them.

3. Cognitive Debt

Metaphor: Paying with a credit card you never read the statement for — the ease now is borrowed against a capability you are quietly spending.

Definition: Using AI to do a cognitive task is not the same as using it to learn one.
When the machine does the thinking, the brain does not light up — and does not grow.
The output looks finished; the learning that output was supposed to cause never happened.
This deficit compounds silently, like debt, until the bill arrives as helplessness.
The danger is greatest exactly where the tool is most seductive: hard, effortful work.
Productive struggle is not an obstacle to learning — it is the learning.

Why it holds:

• The MIT Media Lab EEG study found LLM-assisted writing produced the lowest brain connectivity of any condition, and writers who couldn’t recall their own text — measurable “cognitive debt.”

• A randomised study of “metacognitive laziness” found ChatGPT users gained short-term performance but showed no better knowledge transfer — the learning didn’t stick.

• Wharton’s guardrail experiment showed students given raw GPT to practise with then performed worse on unaided exams than students who practised without it.

• Decades of “desirable difficulties” research (the Bjorks) show that making learning feel harder in the moment is what makes it durable — and AI, misused, removes exactly that difficulty.

• Systematic reviews of AI in higher education report a consistent association between heavy reliance and weaker independent critical thinking.

How to build with it:

• Enforce struggle-first sequencing: students attempt the hard task unaided, then bring AI in to check, extend or critique — never to produce the first draft of their thinking.

• Design tasks where the effort is the point, and make the effort visible and rewarded, not just the output.

• Use AI to increase difficulty where useful — generating harder problems, tougher counter-arguments — rather than to lower it.

• Teach the concept of cognitive debt to students directly; name it, so they can catch themselves taking it on.

• Audit assignments with one test: does this build a capability in the student, or does it let the student rent one? Redesign anything that fails.

4. Metacognitive Laziness

Metaphor: Handing the steering wheel to a chauffeur and then wondering why you never learned the route.

Definition: Metacognition is the mind managing itself — planning, monitoring, correcting.
It is the single most transferable skill in all of education.
A chatbot will happily assume that management role the moment a student lets it.
It plans, it decides what’s relevant, it judges when the work is done — and the student coasts.
The task gets finished, but the self-regulation muscle never fires.
Outsource the driver, and you never become one.

Why it holds:

• The “metacognitive laziness” study found students offloaded their self-regulatory work to ChatGPT, gaining performance without the underlying learning that self-regulation produces.

• Microsoft Research’s survey of knowledge workers found that higher confidence in AI correlated with reduced critical-thinking effort — people stopped monitoring their own reasoning.

• The Education Endowment Foundation identifies metacognition and self-regulated learning as among the highest-impact, best-evidenced, lowest-cost interventions in schooling.

• Zimmerman’s model of self-regulated learning (forethought → performance → self-reflection) is precisely the cycle a chatbot short-circuits when it runs the loop for the student.

• Hattie and Donoghue’s synthesis of 228 meta-analyses places self-regulation strategies among the most powerful levers on achievement.

How to build with it:

• Teach metacognition explicitly and by name — planning, monitoring, self-testing, reflecting — as a core strand of the curriculum, not an afterthought.

• Require students to do the managing even when AI does some of the producing: they set the goal, judge the output, decide when it’s good enough.

• Use AI as a metacognitive coach, not a doer — prompt it to ask the student “what’s your plan?” and “how will you check this?” rather than to hand over answers.

• Build reflection into every project: what did you try, where did you get stuck, what would you do differently.

• Make thinking visible — worked examples, think-alouds, visible reasoning — so students internalise the process the machine would otherwise hide.

5. The Atrophy of Critical Thinking

Metaphor: A guard who trusts every visitor eventually stops checking the badges — and then anyone walks in.

Definition: Critical thinking is the disciplined refusal to accept a claim without grounds.
It is effortful, and effort is exactly what a fluent, confident AI tempts us to skip.
The smoother the machine’s answer, the less inclined we are to interrogate it.
Trust, once habitual, becomes deference; deference becomes the atrophy of judgement.
And AI’s confidence is uncorrelated with its correctness — it is fluent when it is wrong.
A generation that cannot tell can be told anything.

Why it holds:

• Microsoft Research found that greater trust in generative AI predicted less critical evaluation of its outputs — the tool’s confidence displaces the user’s scrutiny.

• Systematic reviews across higher education report that heavier AI reliance is associated with declines in students’ independent critical-thinking dispositions.

• Studies of AI-text detection (Weber-Wulff and colleagues) show even experts and tools struggle to tell machine output from human — so “it sounds right” is a broken heuristic.

• Research on GPT detectors found them biased and unreliable, underscoring that surface fluency carries no signal of truth.

• The Elaboration Likelihood Model (Petty and Cacioppo) explains why: under low effort we accept messages via peripheral cues like fluency, exactly the cue AI maximises.

How to build with it:

• Make verification a non-negotiable step of every AI interaction: students must check, source and challenge what the machine produces before using it.

• Assign adversarial tasks — “find three errors in this AI answer,” “argue the opposite,” “grade the model” — so scrutiny becomes reflexive.

• Teach the epistemics of AI directly: how these systems generate text, why they hallucinate, why confidence is not accuracy (see challenge 22).

• Reward the student who catches the machine, not just the one who uses it smoothly.

• Keep some assessment closed-book and unaided, so the ability to reason without a crutch is built and tested.

6. The Illusion of Knowledge

Metaphor: Standing in a well-stocked library and mistaking the address of the books for the contents of your head.

Definition: Access to information produces a powerful feeling of understanding.
That feeling is very often false.
Retrieving an answer is fast and fluent; building the understanding behind it is slow and hard.
Because the fluent path feels like competence, students stop taking the hard one.
The gap only reveals itself under test, when the source is gone and nothing remains.
Learning has to close the gap between feeling you know and actually knowing.

Why it holds:

• Ward’s experiments show that searching the internet inflates people’s confidence in their own knowledge — they credit the machine’s information to themselves.

• Bjork and colleagues document that students systematically misjudge their own learning, preferring strategies that feel productive (rereading) over ones that work (retrieval).

• Roediger and Karpicke’s test-enhanced learning research shows that the act of retrieval — struggling to produce an answer from memory — is what builds durable knowledge, precisely the step an answer-engine removes.

• Dunlosky’s ranking of study techniques finds the “feels good” methods (highlighting, rereading) near-useless and the effortful ones (practice testing, spacing) most powerful.

• The illusion is amplified by AI because its answers are not just available but articulate — fluency the student borrows and mistakes for their own.

How to build with it:

• Use closed-book explanation as the default proof of learning: if you can explain it from your own head, you know it; if you can only look it up, you don’t.

• Deploy frequent low-stakes retrieval practice — the single most robust technique in the science of learning — to convert the illusion into the real thing.

• Teach students to distrust the feeling of fluency and to calibrate: predict your score, then check it.

• Use “teach-back”: students explain to a peer or to the class, exposing gaps a fluent AI answer would have papered over.

• Sequence AI after first attempting from memory, so the student feels the difference between recall and recognition.

7. Creative Dependence

Metaphor: If you always ask the oracle before you think, you never find out what you would have said.

Definition: Original thought begins in the friction of the blank page.
That friction is uncomfortable, and AI abolishes it on demand.
When the first move is always “ask the model,” the student’s own divergent thinking never fires.
What returns is fluent, plausible, and drawn from the average of everything ever written.
Averaged output is the enemy of originality — it regresses every idea to the mean.
Creativity is a muscle, and a muscle that is never loaded wastes away.

Why it holds:

• A Frontiers in Psychology study links dependence on AI to weaker creative-thinking dispositions among students.

• Generative models are, by construction, engines of the probable — they interpolate the existing corpus, which is the opposite of genuine novelty.

• “Desirable difficulties” research shows that the effortful, uncomfortable phase of a task is where the durable, generative learning happens.

• The MIT cognitive-debt finding extends to ideation: writers who leaned on the model showed less of the neural integration associated with original synthesis.

• Brynjolfsson’s “Turing Trap” argument warns that building AI to imitate rather than augment humans quietly devalues the distinctly human capacities — originality chief among them.

How to build with it:

• Enforce “generate before you generate-with”: students produce their own ideas, sketches or drafts first, and only then use AI to stress-test or extend them.

• Protect the blank page — deliberately un-assisted ideation time — as a scarce and valuable ritual.

• Use AI as a sparring partner for divergence: ask it for ten bad ideas to react against, not one good idea to adopt.

• Reward the idea the machine wouldn’t have produced; grade for surprise, not just polish.

• Teach taste (see Article 3): the judgement to tell a genuinely new idea from a fluent average one.

8. The Unguarded Chatbot

Metaphor: Handing a learner driver a car with the answers to the test taped to the windscreen — they pass, and they cannot drive.

Definition: Not all AI use is equal; the default configuration is often the worst one.
A raw, general chatbot will give the answer because that is what it is built to do.
Given the answer, the student skips the process that would have built the skill.
The result can be measurably negative: worse performance when the crutch is removed.
The harm is invisible in the moment because the homework looks excellent.
An unguarded chatbot is not a neutral tool; it is an active de-skilling agent.

Why it holds:

• Wharton’s randomised study of ~1,000 students found that practising with unguarded GPT lowered subsequent unaided exam scores — and that a hint-only “GPT Tutor” erased the harm.

• The MIT and metacognitive-laziness studies both show the damage flows from the tool doing the cognitive work the learner should be doing.

• Microsoft Research’s finding that reliance dampens critical thinking is strongest where the tool answers directly rather than scaffolds.

• By contrast, the Harvard, Nigeria and Stanford tutor studies — all of which produced large gains — used deliberately designed, pedagogy-first systems, not raw chat.

• The pattern across the evidence is unambiguous: outcome depends on configuration, not on “AI” in the abstract.

How to build with it:

• Ban the raw, general chatbot from the learning core and replace it with guardrailed, pedagogy-first tutors that coach rather than complete.

• Require any classroom AI to withhold final answers by default and to work through hints, questions and steps (see challenge 9).

• Procure and evaluate tools on learning outcomes when the tool is removed, not on how impressive the assisted output looks.

• Teach students the difference so they self-select the right mode even on personal devices.

• Treat “we gave them ChatGPT” as a null strategy — the configuration is the intervention, not the access.

9. The Tutor-Not-Answer-Key Problem

Metaphor: A great coach never plays the match for you; they make you run the drill again, better.

Definition: The central design challenge of AI in education is restraint.
A useful tutor is defined less by what it says than by what it refuses to say.
It must diagnose the misconception, not paper over it with a correct answer.
It must hold back the solution and hand over the next question instead.
This is hard to build, because the model’s instinct is to be maximally helpful — i.e. to tell.
The whole art is engineering an AI that helps by not helping too much.

Why it holds:

• Wharton’s “GPT Tutor,” constrained to give hints rather than answers, neutralised the harm that raw GPT caused — the constraint was the pedagogy.

• Stanford’s Tutor CoPilot, which coaches human tutors in real time rather than replacing them, raised student mastery, with the biggest gains for the weakest tutors.

• The SocraticAI line of work shows LLM tutors can be engineered to enforce dialogue, well-formed questions and usage limits instead of dispensing solutions.

• Bloom’s two-sigma result came from human tutors who diagnosed and scaffolded — the behaviour we now have to encode into software.

• Cognitive-apprenticeship theory (Collins, Brown, Holum) — model, coach, scaffold, fade — is the exact template a good AI tutor should follow.

How to build with it:

• Specify “tutor, not answer-key“ as a hard requirement in every procurement and every custom build: hints, questions and steps by default; answers only after genuine attempts.

• Encode the scaffold-and-fade arc — more support early, deliberately withdrawn as competence grows.

• Have the AI surface and target misconceptions, not just mark right/wrong.

• Keep a human in the loop as the accountable pedagogue; use AI to extend the teacher’s reach, not to remove the teacher (see challenge 11).

• Pilot on the “removed-tool” test: the design is working only if unaided performance improves.

10. The Unclaimed Two-Sigma Prize

Metaphor: For forty years we knew the cure and couldn’t afford the medicine; the price just collapsed.

Definition: In 1984 Benjamin Bloom found one-to-one tutoring lifts the average student two standard deviations.
That is the difference between the middle of the class and the top few per cent.
It was education’s holy grail and its cruelest fact — because tutoring for all was unaffordable.
AI is the first technology with a credible claim to deliver personalised tutoring at scale.
The early randomised trials are not incremental; they are among the largest gains ever measured.
The prize is real — but it is claimed only by schools that deploy the tool as a tutor, not a toy.

Why it holds:

• Bloom’s original two-sigma paper set the benchmark every AI tutor is now measured against.

• The Harvard physics RCT (Kestin and colleagues) found more than double the learning of an active-learning class, in less time, from a purpose-built tutor.

• The World Bank’s Nigeria RCT recorded gains equivalent to roughly two years of schooling from six weeks of AI tutoring — extraordinary cost-effectiveness.

• A meta-analysis of intelligent tutoring systems (Ma and colleagues, 107 effect sizes) found they already outperformed teacher-led and other computer-based instruction before the LLM era.

• Stanford’s Tutor CoPilot shows the gains extend to human tutors augmented by AI, not only to students facing a bot.

How to build with it:

• Treat the two-sigma prize as the north star of adoption: the goal is Bloom’s tutor for every child, finally affordable.

• Pair AI tutoring with mastery pacing — let students move when they’ve mastered a concept, not when the calendar says so.

• Keep the teacher as orchestrator: AI handles personalised practice and feedback; the human handles motivation, judgement and the human relationship.

• Instrument for learning gains, not usage minutes; measure what Bloom measured.

• Prioritise the students who never had access to a tutor — that is where the gains, and the justice, are largest.

11. The Teacher’s New Job

Metaphor: When the printing press arrived, the scribe’s job didn’t vanish — it became the author’s, the editor’s, the publisher’s.

Definition: The teacher-as-lecturer is the role AI most directly disrupts.
Delivering information to thirty passive listeners is the one thing software now does cheaply.
But the teacher’s real job was never information delivery — that was the medium’s limitation.
The real job is diagnosis, motivation, judgement, relationship and the design of experience.
AI doesn’t shrink that job; it removes the drudgery that crowded it out.
The teacher becomes the architect and coach of learning, with a tutor for every student on tap.

Why it holds:

• Stanford’s Tutor CoPilot lifted student outcomes by making tutors better in real time — evidence that the highest-leverage use augments the educator, not replaces them.

• The Nigeria and Harvard gains were realised inside teacher-led programmes; the human set the frame, the AI did personalised practice.

• Deming’s research on the rising labour-market return to social skills implies the human, relational parts of teaching are appreciating, not depreciating.

• The EEF’s evidence base shows the highest-impact moves — feedback, metacognition, relationships — are precisely the human ones AI can support but not supply.

• The US Department of Education’s guidance frames AI as a tool to be kept firmly “in the loop” behind a human educator, not a substitute for one.

How to build with it:

• Retrain teachers as learning architects and coaches: designing tasks, diagnosing misconceptions, mentoring — with AI handling delivery and drill.

• Give every teacher an AI co-pilot for planning, differentiation and feedback, freeing hours now lost to routine production.

• Rewrite the job description and the training: less “cover the syllabus,” more “engineer the conditions for deep learning.”

• Protect and elevate the relational core — the part no model can do — as the profession’s defining value.

• Bring teachers into tool selection and design; the ones who will run these systems must shape them (see challenge 23).

12. The Death of the Take-Home Essay

Metaphor: You cannot test someone’s swimming by asking them to describe a swim they did alone, at home, unwatched.

Definition: The unsupervised written assignment was always a proxy for thinking.
That proxy worked only because producing the artefact required the thinking.
AI severs the link: now the artefact can appear with no thinking behind it.
Trying to police this with detection software is a losing, and unjust, arms race.
The essay isn’t dead as a learning activity — it’s dead as an unsupervised assessment.
Assessment has to move from the product to the process that made it.

Why it holds:

• Weber-Wulff and colleagues tested fourteen AI-text detectors and found them neither accurate nor reliable — the enforcement tool doesn’t work.

• Stanford researchers found GPT detectors systematically misclassify non-native English writers’ work as AI-generated — the tool is not just weak but biased.

• HEPI’s 2025 survey found generative-AI use among UK students had reached ~92%, with the large majority using it for assessment — the practice is already universal.

• TEQSA, QAA and university white papers converge on the same conclusion: redesign assessment, don’t try to detect your way out.

• The University of Pittsburgh’s analysis argues authentic, process-oriented assessment is the only ethical path — surveillance is both ineffective and corrosive of trust.

How to build with it:

• Move the graded moment into the room: oral defences, in-class writing, live problem-solving, presentations, vivas.

• Assess the process, not just the product — drafts, notes, reasoning logs, the trail of how the thinking was built.

• Make AI use explicit and cited where it’s allowed, and design tasks where using it well is itself the skill being assessed.

• Use “flipped” integrity: let students prepare with AI, then demonstrate understanding unaided and in person.

• Retire the unsupervised, un-defended take-home essay as a summative instrument; keep it as low-stakes practice.

13. Integrity Without Surveillance

Metaphor: You don’t stop people cheating at a chess match by installing cameras — you sit them at the board and watch them play.

Definition: The instinct when cheating gets easy is to build a bigger cage.
Detection, plagiarism scanners, lockdown browsers, proctoring spyware — the surveillance reflex.
It fails technically, because the detectors don’t work and the tools evolve faster than the cage.
It fails morally, because it treats every student as a suspect and poisons the relationship.
The durable answer is not to catch cheating but to design it out.
When the only way to complete a task is to learn, cheating and not-learning become the same act.

Why it holds:

• The detector studies (Weber-Wulff; the Stanford bias study) show surveillance-based enforcement is both unreliable and discriminatory.

• The QAA’s guidance for the “ChatGPT era” explicitly recommends programme-level assessment redesign over detection.

• ERIC-indexed studies of academics adapting to AI show a clear migration toward authentic, time-limited, process-oriented tasks that make shortcutting pointless.

• Self-determination theory (Ryan and Deci) predicts that surveillance undermines the intrinsic motivation and trust on which real learning depends.

• Where tasks are personal, oral, iterative or tied to the student’s own context, there is simply nothing generic for a model to hand over.

How to build with it:

• Design assessments that are cheat-proof by construction: personal, oral, in-class, iterative, tied to the student’s own work and context.

• Make the process the deliverable — reasoning trails, drafts, defences — so the learning cannot be skipped.

• Shift from a policing posture to a trust-and-transparency one: agree openly when and how AI may be used, and assess the judgement in using it.

• Invest the money you would have spent on detection software into assessment redesign and teacher time.

• Frame integrity as self-respect — the point of school is to build a mind, and cheating is stealing from yourself.

14. The Equity Fork

Metaphor: The same river can carve a canyon that divides two lands, or irrigate both — direction is a choice.

Definition: AI in education is an amplifier, and amplifiers are neutral about what they amplify.
Left to the market, the best tutors and the best guidance flow to those who already have most.
That path widens every existing gap into a chasm.
But the same technology has a striking property: it helps the least-skilled the most.
Deployed universally and deliberately, it can be the greatest leveller schooling has ever had.
Which fork you take is not decided by the technology; it is decided by policy.

Why it holds:

• Brynjolfsson, Li and Raymond’s field study found generative AI raised productivity most for novice workers, compressing the gap with experts.

• Noy and Zhang found ChatGPT narrowed the performance gap between stronger and weaker writers.

• The World Bank’s Nigeria trial delivered its outsized gains to ordinary secondary students, not an elite — evidence the leveling can be real.

• Stanford’s Tutor CoPilot produced its largest gains for the lowest-rated tutors, again lifting the bottom fastest.

• Conversely, adoption surveys (HEPI; the Digital Education Council) show usage is running ahead of guidance, which without intervention favours the already-advantaged.

How to build with it:

• Treat universal access to high-quality, guardrailed AI tutoring as an equity imperative, funded as infrastructure, not a premium add-on.

• Target the highest-need students first — the leveling potential is concentrated exactly where the need is greatest.

• Standardise the quality of the tool across schools, so the child’s postcode doesn’t determine the tutor’s calibre.

• Pair access with the digital-sovereignty and attention curriculum (challenge 15), so disadvantaged students get the defensive skills too, not just the tool.

• Measure the gap, not just the average; success is the distribution narrowing.

15. The Attention-Economy Adversary

Metaphor: You are trying to teach in a casino that has been engineered, floor to ceiling, to make sure nobody ever leaves the slot machines.

Definition: School does not compete for attention on a level field.
It competes against the most sophisticated persuasion machinery ever built.
Billions of dollars and the best behavioural science are spent to capture the same minutes learning needs.
The feed is not a distraction that happens to exist; it is an adversary optimised to win.
Its business model is literally the conversion of attention into revenue — the exact resource education requires.
You cannot ignore an adversary this good; you have to name it and out-design it.

Why it holds:

• James Williams’ Stand Out of Our Light argues the attention economy is structurally at war with human intention and self-determination.

• Research on dark patterns (Mathur and colleagues; the NSF taxonomy) catalogues how interfaces are deliberately engineered to manipulate behaviour against the user’s interest.

• The OECD’s report on dark commercial patterns documents their prevalence and measurable effectiveness at scale.

• The Norwegian Consumer Council’s Deceived by Design shows platforms steering users away from their own privacy and autonomy through design.

• CIGI’s work on the teen brain shows adolescent reward and attention systems are specifically the ones this machinery is optimised to exploit.

How to build with it:

• Teach the adversary directly: a digital-sovereignty curriculum on persuasive design, dark patterns and the attention economy, so students can see the hooks.

• Shrink the surface area during the learning core — phone-free defaults, single-purpose devices, notification-free environments.

• Make attention training (challenge 1) explicit countermeasures: the ability to reclaim focus is now a defensive life skill.

• Model and teach deliberate technology use — intention before device — as a habit, not a rule.

• Frame the goal as sovereignty over one’s own mind, the master-skill of the century, and sell it to students as power, not restriction.

16. The Motivation Problem

Metaphor: Why climb the mountain when a helicopter will drop you on the summit — and hand you a photo to prove you were there?

Definition: Effort has always been sustained by the sense that it was necessary and yours.
AI quietly removes the necessity: any answer is one prompt away.
If the mountain can be skipped, the will to climb it collapses.
And extrinsic motivators — grades, compliance — are exactly what AI makes easiest to game.
The only motivation that survives is intrinsic: autonomy, mastery, purpose, real stakes.
An AI-era school has to rebuild its motivational engine on foundations the machine cannot counterfeit.

Why it holds:

• Self-determination theory (Ryan and Deci) identifies autonomy, competence and relatedness as the roots of durable, intrinsic motivation — none of which a shortcut satisfies.

• The metacognitive-laziness and cognitive-debt findings show that when the extrinsic goal (finish the task) can be met without effort, the effort — and the learning — evaporates.

• Duckworth’s work on grit ties long-term achievement to sustained passion and perseverance toward goals the person actually owns.

• OECD’s “Student Agency for 2030” frames agency and ownership of learning as central design goals precisely because compliance-based motivation is failing.

• The exemplar models (High Tech High, Montessori, microschools) that sustain motivation do so through real projects, autonomy and mastery — not through easier extrinsic rewards.

How to build with it:

• Rebuild motivation on autonomy, mastery and purpose: give students genuine choice, visible progress toward competence, and reasons that matter to them.

• Anchor work in real stakes — authentic audiences, real projects, real consequences — so completion is meaningful, not performative (see challenge 24 and Article 3’s “Playing for Real”).

• Use mastery pacing so competence is felt and earned, feeding the intrinsic loop.

• Make relationships central; the relational bond with a teacher and peers is a motivational force AI cannot supply.

• Design out the gameable extrinsic reward wherever it dominates; reward depth, effort and growth instead.

17. The Meaning Problem

Metaphor: If a robot can lift the weights for you, the gym only makes sense once you realise you came to build your own body.

Definition: There is a question every AI-era student will eventually ask out loud.
If the machine can write it, solve it, make it — why should I learn to?
Answered badly, it is corrosive: it makes all of school look pointless.
Answered well, it is liberating: because learning was never only about producing the output.
Learning is self-formation — it is the process by which a person becomes capable, and becomes themselves.
The output was always a by-product; the real product is the mind and the character that did it.

Why it holds:

• Damon’s research on the development of purpose shows that a sense of meaning is a powerful driver of engagement, wellbeing and persistence in young people.

• Seligman and Adler’s positive-education work makes the case that wellbeing and flourishing belong in the curriculum as explicit aims, not accidents.

• Self-determination theory ties motivation to purpose and relatedness — a “why” that survives the availability of shortcuts.

• The future-of-work evidence (WEF; Autor) implies that as machines do more doing, the human premium shifts to judgement, direction and meaning-making — capacities formed through effortful learning.

• Character-education frameworks (the Jubilee Centre; Oxford) argue that education’s oldest purpose — forming a person — is exactly the purpose AI cannot outsource.

How to build with it:

• Reframe the purpose of learning explicitly and often: you are not producing essays, you are building yourself — a mind, a character, a set of powers no one can take.

• Put purpose, meaning and character into the curriculum as first-class strands (see Article 3’s “Why Anything At All”).

• Connect learning to the student’s own goals and to real contribution, so the “why” is felt, not lectured.

• Distinguish for students between tasks (delegate freely) and formation (never delegate) — the machine can do your work, not your becoming.

• Treat meaning as the antidote to the shortcut: a student who knows why they climb does not want the helicopter.

18. The Knowledge-Is-Obsolete Fallacy

Metaphor: You cannot connect the dots you don’t have; “just look it up” gives you a screen full of dots and no lines.

Definition: The seductive error of the AI age is that facts no longer matter.
Why memorise anything, the argument runs, when everything is instantly retrievable?
Cognitive science says this is precisely backwards.
Thinking, comprehension and creativity all run on knowledge held in the mind, not on tap.
You cannot think critically about a subject you know nothing about — critical thinking is domain-specific.
The more the machine can retrieve, the more valuable it becomes to be the human who actually knows.

Why it holds:

• Willingham’s cognitive-science work shows critical thinking is not a free-floating skill but is bound to deep domain knowledge — you reason well only about what you understand.

• Kirschner, Sweller and Clark demonstrate that reasoning happens in working memory drawing on richly organised long-term knowledge; skills-without-knowledge pedagogies underperform.

• The National Research Council’s Education for Life and Work concludes transferable competencies develop through rich content, not instead of it.

• Hirsch-lineage critiques (Rotherham and Willingham) show “21st-century skills” fail when divorced from a knowledge-rich curriculum.

• The memory-paradox and offloading research (challenge 2) confirms that offloaded knowledge doesn’t build the schemas comprehension requires.

How to build with it:

• Keep building deep, structured domain knowledge — AI raises the value of the knowledgeable human, it doesn’t remove the need for one.

• Reject the “skills, not facts” false binary; teach powerful knowledge and the skills that only fluent knowledge makes possible.

• Use AI to deepen knowledge — richer examples, faster feedback, more practice — not to excuse its absence.

• Sequence carefully: build the schema first, then let AI extend reach; never let retrieval substitute for understanding.

• Treat a knowledge-rich curriculum as the precondition for everything else in this list — you cannot verify, create or judge from an empty head.

19. The Depth Tradeoff

Metaphor: A machine that lets you fill a hundred shallow holes will always tempt you away from digging one deep well.

Definition: AI makes breadth and speed almost free.
It makes it trivially easy to complete a great deal, quickly and passably.
Completion, though, is not the same as depth, and school confuses the two at its peril.
Depth comes from revision, from wrestling, from returning to a hard thing until it yields.
The frictionless path AI opens leads straight past exactly that.
An AI-era school has to actively re-price depth above the shallow completion the tool rewards.

Why it holds:

• The cognitive-debt and desirable-difficulties evidence shows durable understanding comes from effortful depth, not frictionless breadth.

• Bloom’s mastery tradition and competency-based education both insist on depth-to-mastery over coverage-for-its-own-sake.

• Deliberate-practice research (Ericsson) locates expertise in sustained, focused work at the edge of ability — the opposite of skimming many tasks.

• Deeper-learning evaluations (AIR) associate depth-oriented models with higher graduation and college enrolment, not just test scores.

• Project-based and mastery models (MDRC’s PBL review; RAND on personalised learning) show depth-oriented design produces stronger outcomes when done well.

How to build with it:

• Re-price the reward: grade and celebrate depth, revision and defence of ideas, not the volume or polish of what was completed.

• Assign fewer, deeper tasks — one well dug beats a hundred holes — and give time to return and improve.

• Require iteration: multiple drafts, critique, revision, with the improvement itself assessed.

• Make students defend their work in person, where depth (or its absence) is immediately visible.

• Use AI to enable depth — more feedback cycles, harder challenges — rather than to multiply shallow output.

20. The Developmental Mismatch

Metaphor: You don’t give a learner a forklift before they’ve built the muscles that tell them what a heavy thing feels like.

Definition: A student is not a small adult with less information.
A child’s brain is still building the very machinery AI offers to bypass.
Executive function, working memory and self-regulation are laid down through effortful use, over years.
Hand a still-forming mind a tool that does the effortful part, and the machinery may never fully build.
What is a reasonable delegation for a skilled adult can be a developmental theft for a child.
Age and stage have to govern how, and how much, AI enters the picture.

Why it holds:

• Diamond’s synthesis shows executive functions develop through practice across childhood and adolescence — they are built, not issued.

• Harvard’s Center on the Developing Child describes executive function as an “air-traffic-control system” constructed through early, effortful experience.

• Blakemore, Steinberg and Fuhrmann show the adolescent brain is a sensitive period of heightened plasticity — and of an immature control system paired with a strong reward system.

• The offloading research (challenge 2) implies that offloading during the formative window risks skipping the construction of the underlying capacity.

• The screen-time and adolescent-brain evidence (Pew; the Frontiers scoping review; CIGI) shows developing minds are uniquely vulnerable to attention-hostile technology.

How to build with it:

• Age-gate AI deliberately: the youngest children build fundamentals — reading, number, executive function, focus — before general AI enters the picture.

• Sequence delegation to development: as the underlying capacity is built and demonstrated, expand what may be offloaded.

• Keep the effortful, capacity-building work un-automated during the windows when that capacity is being laid down.

• Match tool design to stage — heavily scaffolded and bounded for the young, more open for the older and more capable.

• Make “build the muscle before you use the machine” an explicit, stage-based policy, not an ad-hoc classroom call.

21. Data, Privacy and the Student Profile

Metaphor: A tutor who remembers everything a child ever struggled with is a gift; the same memory in the wrong hands is a dossier.

Definition: Personalised AI works by knowing the learner in detail.
That intimacy is the source of its power — and of a serious new risk.
Children’s learning data is uniquely sensitive: their struggles, their pace, their private patterns.
Aggregated over years, it becomes a profile more revealing than any report card.
The attention-economy players building these tools have business models built on exactly such data.
Protecting the learner’s data is now inseparable from protecting the learner.

Why it holds:

• The dark-patterns and Deceived by Design research shows how routinely platforms engineer users away from protecting their own data.

• UNESCO’s and the OECD’s guidance on generative AI in education both foreground data protection, age limits and privacy safeguards as first-order concerns.

• The US Department of Education’s AI guidance stresses privacy, transparency and keeping humans accountable for automated decisions about students.

• CIGI’s work on the teen brain underlines how design optimised for engagement exploits exactly the population schools are meant to protect.

• The concentration of these tools among a few commercial actors raises the stakes of who holds, and can monetise, the student profile.

How to build with it:

• Set data governance as a precondition of adoption: know what is collected, where it lives, who can see it, and for how long — or don’t deploy.

• Prefer tools with data-minimisation and strong privacy guarantees; treat student data as a liability to be minimised, not an asset to be hoarded.

• Keep a human accountable for any consequential decision an AI system informs about a student.

• Teach students their own data rights and digital footprint as part of the sovereignty curriculum (challenge 15).

• Make privacy a procurement gate, not an afterthought — the intimacy that powers the tutor is exactly what must be protected.

22. The Hallucination Problem

Metaphor: A brilliant, tireless assistant who is also a fluent, unembarrassed liar — and never once tells you which it’s being.

Definition: Large language models do not know things; they predict plausible text.
Most of the time the plausible text is also true, which is what makes the failures dangerous.
When they are wrong, they are wrong with exactly the same fluent confidence as when they are right.
There is no tremor in the voice, no hedge, no tell.
A student who trusts the surface will absorb falsehoods as readily as facts.
So verification cannot be optional; it has to become a reflex, trained until automatic.

Why it holds:

• The detector research (Weber-Wulff; the bias study) confirms fluency carries no signal of truth — you cannot tell right from wrong by how it reads.

• Microsoft Research found that trust in AI displaces the critical scrutiny that would catch its errors.

• The Elaboration Likelihood Model explains why confident fluency is persuasive under low effort — the precise condition of a rushed student.

• Systematic reviews link heavy AI reliance to weaker independent judgement — the muscle needed to catch hallucination.

• The epistemics and verification literature (see Article 3’s “Discernment”) establishes calibration and source-checking as trainable skills.

How to build with it:

• Teach distrust-and-verify as a reflex: never use an AI claim you haven’t checked against a real source.

• Make how models work explicit — prediction, not knowledge — so students understand why they must verify.

• Build verification drills into everyday AI use: cross-check, cite, catch the error, rate the confidence.

• Reward the student who spots the hallucination; make catching the machine a graded, celebrated skill.

• Keep verification tied to a knowledge-rich curriculum (challenge 18) — you can only catch a falsehood about something you actually understand.

23. The Change-Management Wall

Metaphor: You can install a new engine overnight, but the crew that has to run it turns over only at the speed of trust.

Definition: The hardest part of AI in education is not the AI.
It is the human system — teachers, parents, institutions — that must actually change.
Mandates from above produce compliance, resentment and quiet sabotage, not transformation.
Teachers who don’t trust or understand a tool will not use it well, and rightly so.
Change that lasts is led by educators, grounded in evidence, and started where the leverage is highest.
Ignore the change-management problem and the best technology in the world sits unused.

Why it holds:

• Stanford’s Tutor CoPilot succeeded by supporting educators in their existing work — change that ran with the grain of the profession, not against it.

• The EEF’s implementation evidence shows that how an intervention is adopted matters as much as what it is.

• History (challenge 26 in the library: the persistence of the industrial school) shows education systems are extraordinarily resistant to structural change.

• Adoption surveys reveal a governance vacuum — students racing ahead while institutions lag — a classic failure of managed change.

• Self-determination theory applies to teachers too: autonomy and competence drive genuine adoption; coercion drives the opposite.

How to build with it:

• Lead with teachers, not mandates: involve educators in selection and design, and build their competence and confidence first.

• Start with the highest-leverage, lowest-risk uses — teacher planning, feedback, differentiation — and let visible wins build trust.

• Run evidence-led pilots, measure learning outcomes, and scale what works rather than what is fashionable.

• Bring parents and students into the frame with a clear, honest account of the two mandates — adopt and defend.

• Treat adoption as a multi-year cultural change, resourced and led as such — not a procurement event.

24. The Cost of Not Adopting

Metaphor: Standing still on a moving walkway feels like safety; it is just a slower way of being carried somewhere you didn’t choose.

Definition: The safest-sounding option is to keep AI out and carry on.
It is also, quietly, the most expensive.
Refusing to adopt is not neutrality; it is a decision to prepare students for a vanished world.
The labour market is being rewritten around people who can direct these tools.
A graduate who has never learned to work with AI enters that market already behind.
Not adopting is a choice — and it is the costlier one, borne by the students who had no say in it.

Why it holds:

• The WEF’s employers project 170 million new roles and 92 million displaced by 2030, with AI-and-data fluency among the fastest-rising skills.

• Eloundou and colleagues estimate around 80% of US workers have at least some tasks exposed to LLMs — the technology touches nearly every future job.

• Autor’s work argues AI, used well, can rebuild middle-skill work by extending expertise to more people — an opportunity available only to those trained to seize it.

• Brynjolfsson, Li and Raymond, and Noy and Zhang, show AI most helps those who learn to use it — the skill is learnable, and its absence is a handicap.

• Deming’s evidence on the rising return to human-plus-technology skills implies the cost of exclusion compounds over a career.

How to build with it:

• Reframe the risk honestly: the question is not “is adopting risky?” but “which risk do we choose — the manageable risks of adopting well, or the unmanaged risk of not adopting at all?”

• Adopt deliberately, not defensively: teach with AI and about AI, aimed squarely at the capacities that appreciate.

• Make AI fluency — orchestration, verification, collaboration — an explicit learning goal (see Article 3’s “Commanding the Swarm”).

• Move now, but move designed: every challenge in this article is a reason to adopt carefully, none is a reason to abstain.

• Own the decision: choosing not to prepare students for their actual future is the one choice with no defence.

So what: two mandates, one classroom

Read the twenty-four together and a single architecture emerges. Every challenge is a variation on one theme, and every fix is a variation on one move. The theme is that artificial intelligence will either amplify a mind or replace the work that builds one — and it does whichever the design tells it to. The move is to run two mandates at once, deliberately, in the same room, on the same day: teach fluently with AI, and defend the mind from it.

Concretely, that means a school built on a small number of non-negotiables. Protect attention as the master resource and teach it as a subject. Insist on struggle-first sequencing, so the machine amplifies effort instead of erasing it. Keep the load-bearing knowledge inside the student’s own head, and make closed-book explanation the proof of learning. Replace the raw chatbot with guardrailed, Socratic tutors that coach rather than complete — and chase Bloom’s two-sigma prize with them, for every child, starting with those who never had a tutor. Move assessment into the room and onto the process, and abandon both the take-home essay and the surveillance that tried to save it. Teach the attention economy as the adversary it is. Rebuild motivation and meaning on foundations the machine cannot counterfeit. Govern the data. Train verification until it is a reflex. Lead the change through teachers, not over them. And adopt — deliberately, urgently — because the one option with no defence is preparing children for a world that will not exist.

None of this is exotic. Almost every fix in this article is something great educators have always done — diagnose, scaffold, demand depth, build relationships, form character — now made both more necessary and, with the right tools, more achievable than ever before. Artificial intelligence does not change the goal of education. It raises the stakes on getting it right, and it hands us, for the first time, a tutor for every child capable of helping us get there. The schools that thrive will not be the ones that ban the future or the ones that surrender to it. They will be the ones that hold both truths at once: bring the machine all the way in — and never, for a moment, stop defending the mind it could either build or hollow out.

Built on a purpose-built ENSI research library of 179 primary documents across 30 research angles — from the neuroscience of attention and memory to the economics of the agentic labour market — spanning the OECD, UNESCO, the World Economic Forum, the World Bank, the MIT Media Lab, Harvard, Stanford, NBER, the Education Endowment Foundation and dozens more. Part one of the ENSI “Education for the Agentic Age” series.

There is a sentence of Emerson’s that almost everyone has met and almost no one has read. “Trust thyself: every heart vibrates to that iron string.” It arrives now on gym walls and graduation mugs and the closing slide of motivational decks, sanded down into a pat on the back — believe in yourself, you’ve got this. That is not what Emerson wrote, and the distance between the mug and the essay is the whole subject of this piece. The mug says your feelings are valid. The essay says something far stranger and far more demanding: that the voice you hear in solitude is not merely yours, that it is continuous with the intelligence that built the planet and hung the stars, and that to betray it for the good opinion of a room full of people is not a social lapse but a kind of blasphemy against the only thing in you that is real.

Self-Reliance, published in the Essays: First Series of 1841, is the most influential essay an American ever wrote, and it is routinely misfiled. It gets shelved under motivation, under rugged individualism, under the American cult of the self-made man — as if Emerson were an early life coach handing out permission slips. Read the actual paragraphs and a different document appears. It is not a pep talk; it is an epistemology, an ethics and a theory of power welded together. Its claim is not that you should feel good about yourself. Its claim is that conformity is a cognitive failure — that when you defer to custom, to party, to consistency, to the crowd, you are not just being timid, you are seeing falsely, your “every truth is not quite true, your two is not the real two, their four not the real four.” Self-reliance, for Emerson, is first of all a way of knowing, and only afterwards a way of behaving.

That is why the essay has outlived its century while the self-help it spawned curdles every decade. The self-help reading collapses because it keeps the confidence and throws away the metaphysics — and without the metaphysics, “trust yourself” is just narcissism with better lighting. Emerson knew this. He built the whole doctrine on a foundation that has nothing to do with self-esteem: the conviction that there is a single deep source of insight, which he called by many names — the Over-Soul, the aboriginal Self, Spontaneity, Instinct, Intuition — and that each individual mind is a channel to it. Self-trust is warranted not because you are special but because you are connected. “We lie in the lap of immense intelligence, which makes us receivers of its truth and organs of its activity.” Strip that sentence out and the doctrine dies. Keep it, and self-reliance becomes what Emerson meant it to be: the discipline of clearing away everything secondhand — inherited opinion, borrowed taste, the reflex of agreement — so that something first-hand can get through.

This matters now for reasons Emerson could not have foreseen and would have relished. We live in the most powerful conformity engine ever built. The nineteenth-century “joint-stock company” of society that Emerson described — the arrangement whereby “the members agree, for the better securing of his bread to each shareholder, to surrender the liberty and culture of the eater” — is a quaint machine next to a recommendation algorithm that has metabolised the attention of three billion people and learned, to the millisecond, how to make you want what everyone else wants. When Emerson wrote that “society everywhere is in conspiracy against the manhood of every one of its members,” he was describing a village. We have industrialised it. The essay reads today less like a period piece and more like a warning that arrived a hundred and eighty years early.

And it matters for a second reason, closer to ENSI’s own preoccupations. We are entering an age in which cognition itself is delegable — in which the reflex of “what does the model say?” threatens to become the new “what does the party say?”, a fresh and frictionless source of borrowed thought. Emerson’s essay is, among other things, the sharpest instrument we have for telling the difference between a tool that extends your judgement and a habit that replaces it. He drew that line in 1841, against books and traditions and sages. The line is the same. Only the tempter has changed.

So this is a reconstruction, not a quotation-fest. Below are twelve principles — the load-bearing beams of Emerson’s philosophy of self-confidence, drawn from Self-Reliance itself and cross-checked against the essays that surround it (The Over-Soul, Compensation, Circles, Spiritual Laws), the later and colder wisdom of The Conduct of Life, and the century of critics who have tried, mostly in vain, to catch him in contradiction. Read them as a doctrine, because that is what they are. And notice, as you go, how few of them are about confidence at all.

The main points, in brief:

• Self-reliance is an epistemology before it is an attitude. Emerson’s first claim is about truth, not feelings: the thought you dismiss because it is merely yours is the one worth having, and conformity makes you see the world falsely.

• It rests on the Over-Soul, not the ego. Self-trust is licensed because the individual mind is a channel to a universal intelligence (”we lie in the lap of immense intelligence”). Remove that foundation and the doctrine becomes the narcissism it is so often mistaken for.

• Conformity is the great enemy — and it is a form of cowardice and of blindness at once. “Whoso would be a man must be a nonconformist.” Society is a “joint-stock company” that trades your originality for your safety.

• Consistency is the second enemy. “A foolish consistency is the hobgoblin of little minds.” The demand that you agree with your past self is a cage; live in the “thousand-eyed present.”

• The self you can rely on is built, not found. Character is “cumulative”; power lives “in the moment of transition,” in the soul that becomes. Self-reliance is a practice with compound interest, not a mood.

• Imitation is suicide; envy is ignorance. Your only real contribution is the thing only you can do. “Insist on yourself; never imitate.”

• Do your work — vocation is the form self-trust takes in the world. Each person has “that plot of ground which is given to him to till”; strength appears only in the exercise.

• Nothing external can save you — and nothing external can finally harm you. “Nothing can bring you peace but yourself.” Reliance on property, reputation, travel and institutions is the want of self-reliance.

• Compensation governs everything — deal with cause and effect, not with luck. The moral universe is double-entry bookkeeping; there is no escaping the ledger and no need to.

• The doctrine has a cost, and Emerson names it. It “demands something godlike,” it can read as cold, and it was fired at real reformers. The tension between self-reliance and solidarity is real, not a misreading.

• Its afterlife proves the point — Nietzsche, the pragmatists, the psychology of self-efficacy, and the modern self-help industry all descend from it, and each kept a different half.

• Its final principle is antifragile self-trust: “Nothing can bring you peace but the triumph of principles” — peace comes not from outcomes but from alignment with a law you did not invent.

How to read this doctrine

A warning about method, because Emerson invites the wrong kind of reading. He wrote in flashing aphorisms, each sentence polished to detonate on its own, and the essays do not argue in straight lines — they spiral, circle back, contradict, and dare you to complain. This is deliberate. A man who has just told you that “a foolish consistency is the hobgoblin of little minds” is not going to build you a syllogism. The critics who accuse him of incoherence — and they have, from John Morley’s skeptical Victorian appraisal to Paul Elmer More’s New-Humanist reassessment, both in this library — are half right and wholly missing the point. Emerson thought in what he called circles: every truth is a provisional horizon that a larger truth will encircle tomorrow. To systematise him is to falsify him a little. But to refuse to systematise him is to leave the reader with a bag of glittering fragments and no doctrine — which is exactly how the self-help industry got hold of him.

So the twelve principles below are a scaffold, not a cage. They are ordered to build: the first three establish what self-reliance is and what it is grounded in; the middle six work out its consequences for how you think, work and live; the last three confront its costs and its afterlife. Read across them and a single argument emerges — that self-confidence, properly understood, is not a feeling you generate about yourself but a relation you maintain to a source of truth that runs through you, and that almost everything society trains you to do severs that relation. Now the beams.

1. Trust thyself — because the thought you reject is the one worth having

Emerson opens not with the self but with a peculiar observation about reading. “I read the other day some verses written by an eminent painter which were original and not conventional. The soul always hears an admonition in such lines.” The lesson he draws is the essay’s first principle and its seed: “To believe your own thought, to believe that what is true for you in your private heart is true for all men — that is genius.”

Notice the structure of the claim. It is not believe your thought because it is yours (that would be mere assertion). It is: the thought that is true in your private heart is true universally — the private and the universal are the same thing seen from two angles. “Speak your latent conviction, and it shall be the universal sense; for the inmost in due time becomes the outmost.” This is why he can say, without arrogance, that “the highest merit we ascribe to Moses, Plato, and Milton is, that they set at naught books and traditions, and spoke not what men, but what they thought.” Genius is not superior equipment. Genius is the refusal to launder your own perception through other people’s.

And here is the sting, the observation that turns a platitude into a discipline: “In every work of genius we recognize our own rejected thoughts; they come back to us with a certain alienated majesty.” You have had the thought. You dismissed it because it was yours — too familiar, too unbacked by authority — and then a braver mind published it and you called it genius. “Else, tomorrow a stranger will say with masterly good sense precisely what we have thought and felt all the time, and we shall be forced to take with shame our own opinion from another.” Self-reliance begins as the plain refusal to keep doing this: to stop outsourcing the licensing of your own perceptions to whoever says them louder. That is not confidence as a feeling. It is confidence as an epistemic policy — a standing decision about whose reports of reality you will treat as admissible evidence, starting with your own.

2. The Over-Soul — why self-trust is not self-worship

This is the principle the mug throws away, and without it the whole edifice is unstable, so it comes second by right. Ask the hard question the essay itself asks: why is self-trust warranted? What entitles anyone to treat their private perception as a window onto universal truth rather than as mere opinion, prejudice, appetite? Emerson answers it directly, in the essay’s metaphysical core: “The inquiry leads us to that source, at once the essence of genius, of virtue, and of life, which we call Spontaneity or Instinct… We denote this primary wisdom as Intuition, whilst all later teachings are tuitions.”

The self you are told to rely on is not your ego, your preferences, your personality — Emerson has withering contempt for all of that. It is what he calls “the aboriginal Self, on which a universal reliance may be grounded.” Beneath the chattering, willing, opinionated surface self there is a deeper receiver, and what it receives is not generated by you: “We lie in the lap of immense intelligence, which makes us receivers of its truth and organs of its activity. When we discern justice, when we discern truth, we do nothing of ourselves, but allow a passage to its beams.” He develops the same idea at book length in the companion essay The Over-Soul (in this library both in the Gutenberg First Series and as a standalone EmersonCentral text): “that Unity, that Over-Soul, within which every man’s particular being is contained and made one with all other.”

The consequence is decisive and almost always missed. Self-reliance is God-reliance in disguise. “To talk of reliance is a poor external way of speaking. Speak rather of that which relies, because it works and is.” The thing you rely on is not you; it is what works through you when you get your secondhand opinions out of the way. This is why Emerson can be simultaneously the prophet of the sovereign individual and a man with no patience for egotism: the “self” of self-reliance is the least personal thing about you. The scholarship in this library that takes the doctrine seriously — O’Dwyer’s reading of self-reliance as a factor in a flourishing life, the IJWOS study of “integrity and intuition as the heart of self-reliance,” Kovalainen on Emersonian moral perfectionism — all circle this point: that Emerson’s individualism is underwritten by a metaphysics of connection, and that severing the two produces a monster he would not have recognised. The modern self-help that kept the “trust yourself” and dropped the Over-Soul is not a simplification of Emerson. It is his exact inversion.

3. Nonconformity — the joint-stock conspiracy against your mind

Having established what self-trust is (an epistemic policy) and why it is licensed (the Over-Soul), Emerson names the enemy, and names it without mercy. “Society everywhere is in conspiracy against the manhood of every one of its members. Society is a joint-stock company, in which the members agree, for the better securing of his bread to each shareholder, to surrender the liberty and culture of the eater. The virtue in most request is conformity. Self-reliance is its aversion.”

Read that as the systems description it is. Society offers a trade — security in exchange for originality — and the trade is rational for each shareholder, which is exactly why it is so hard to refuse and so corrosive in aggregate. You surrender your independent judgement and in return you get bread, belonging, the absence of “the world’s displeasure.” Emerson’s verdict is total: “Whoso would be a man must be a nonconformist. He who would gather immortal palms must not be hindered by the name of goodness, but must explore if it be goodness. Nothing is at last sacred but the integrity of your own mind.”

But — and this is where the careless reader goes wrong — nonconformity is not contrarianism. Emerson is not telling you to be difficult, to reject things because they are accepted; that would just be conformity with a minus sign, still letting the crowd set your agenda. “I am ashamed to think how easily we capitulate to badges and names, to large societies and dead institutions.” The target is the capitulation, the reflex of deference, whether it points toward agreement or rebellion. And he is clear-eyed about the price. “For nonconformity the world whips you with its displeasure.” He even anatomises the physiology of the pressure — “the foolish face of praise, the forced smile which we put on in company where we do not feel at ease” — the small muscular betrayals by which we sell ourselves for approval. The essay’s most quoted line about being misunderstood belongs here, and it is usually mutilated into mere reassurance. Emerson’s point is not that being misunderstood is fine. It is that originality guarantees misunderstanding, because the crowd computes your orbit from your past and the original act has no past: “To be great is to be misunderstood.” That is not consolation. It is a job description.

4. Against consistency — the right to contradict yourself

The second enemy is subtler than the crowd, because it lives inside you: your own past. “The other terror that scares us from self-trust is our consistency; a reverence for our past act or word, because the eyes of others have no other data for computing our orbit than our past acts, and we are loth to disappoint them.”

Here is the essay’s most famous sentence, and — like “trust thyself” — one of its most misread: “A foolish consistency is the hobgoblin of little minds, adored by little statesmen and philosophers and divines.” The adjective foolish is load-bearing and almost always dropped. Emerson is not against all consistency; he is against the consistency that is foolish — the kind that binds you to a position because you once held it, that makes you defend yesterday’s error to protect your reputation for reliability. “Speak what you think now in hard words, and tomorrow speak what tomorrow thinks in hard words again, though it contradict everything you said today.”

The deep principle underneath is temporal, and it connects self-reliance to Emerson’s whole metaphysics of flux (worked out in Circles, also in this library): live in the present tense of your own mind. “Bring the past for judgment into the thousand-eyed present, and live ever in a new day.” The person who cannot contradict himself is dragging “this corpse of your memory” behind him, letting a dead self veto a living one. And Emerson offers a genuine consolation for the fear that inconsistency means incoherence — the promise that an honest life has a hidden unity you needn’t engineer: “The voyage of the best ship is a zigzag line of a hundred tacks. See the line from a sufficient distance, and it straightens itself to the average tendency.” You do not need to manufacture consistency by clinging to old positions; if you are honest in each moment, the coherence takes care of itself, visible only from a height. This is the principle that most directly rebuts the “incoherence” charge his critics love: Emerson contradicts himself on purpose, as a matter of doctrine, because he trusts the deeper unity of an honest mind more than the surface unity of a defended one.

5. Character is cumulative — the self is built, not found

If principles 1–4 are about clearing away the false, principle 5 is the one that saves the doctrine from vapidity — because it insists that the self you rely on is not simply there for the trusting. It is earned. “The force of character is cumulative. All the foregone days of virtue work their health into this.”

This is the beam that the “believe in yourself” reading cannot even see, and its absence is why the self-help version rings hollow. Emerson is not saying: trust the self you happen to have. He is saying: there is a self worth trusting, and it is accreted, day by day, out of “a train of great days and victories behind.” The magnetism of a great figure “in the senate and the field” is “the consciousness of a train of great days and victories behind. They shed an united light on the advancing actor.” Self-reliance is therefore not a starting condition but a result; it has compound interest. You become reliable-to-yourself by a long series of acts in which you backed your own perception and it held.

And its engine is not stasis but motion — this is the metaphysical heart of the essay and the sentence most worth memorising: “Life only avails, not the having lived. Power ceases in the instant of repose; it resides in the moment of transition from a past to a new state, in the shooting of the gulf, in the darting to an aim.” The self you can rely on is not a noun but a verb: “the soul becomes.” This is why Emerson has such contempt for resting on past achievement, inherited status, accumulated reputation — all of it is “having lived,” dead capital. The living self is only ever the one currently in transition, currently choosing. Self-confidence, in this light, is not a store you draw down; it is a current you either keep flowing or lose. The scholarship on Emerson’s ethics of self-cultivation (Rao’s study of his individualism, the readings of “Emersonian moral perfectionism” in the library) all fasten on this: that the “self” is a project of continuous self-overcoming, which is precisely the strand Nietzsche would carry off and radicalise.

6. Insist on yourself; never imitate

From the metaphysics of the becoming self, Emerson draws a sharp practical rule — the one aimed straight at the anxious and the derivative. “Insist on yourself; never imitate. Your own gift you can present every moment with the cumulative force of a whole life’s cultivation; but of the adopted talent of another, you have only an extemporaneous, half possession.”

The argument is not aesthetic preference; it is a claim about leverage. Your own gift compounds (principle 5); a borrowed one does not, because you did not build the years of cultivation behind it. “That which each can do best, none but his Maker can teach him… Where is the master who could have taught Shakespeare? Where is the master who could have instructed Franklin, or Washington, or Bacon, or Newton? Every great man is a unique.” The imitator is trying to draw down capital he never deposited. This is also where the essay’s earlier hammer-blow lands with full force — “envy is ignorance… imitation is suicide” — the two crimes named in a single breath near the opening and resolved here. Envy is ignorance because it misunderstands the structure of value: your one face, one character, one fact is “not without preëstablished harmony,” placed exactly where one ray of the universe should fall, “that it might testify of that particular ray.” To envy another’s gift is to be ignorant of the specific testimony only you can give. Imitation is suicide because in copying another you kill the one contribution the universe was waiting on from you and you alone. “We but half express ourselves, and are ashamed of that divine idea which each of us represents.”

7. Do your work — vocation as the body of self-trust

Self-reliance would be an abstraction — a mood in a hammock — if it did not touch the ground, and Emerson makes sure it does. It touches ground in work. This is the principle that most surprises readers who expect a mystic: Emerson is relentlessly practical about labour. “The power which resides in him is new in nature, and none but he knows what that is which he can do, nor does he know until he has tried.”

That last clause is the whole ethic of vocation in miniature: you cannot know your own power in advance; it is disclosed only in the exercise. Self-trust is not introspective confidence; it is empirical, discovered by acting. Hence the recurring imperative — “But do your work, and I shall know you. Do your work, and you shall reinforce yourself” — where work is both how you become legible to the world and how you compound the self of principle 5. Each person has “that plot of ground which is given to him to till”; the good corn “can come to him but through his toil bestowed” on it. Emerson’s late book The Conduct of Life (in this library in three copies, including the 1860 first edition) hardens this into a philosophy of power and practical energy, but the seed is here: self-reliance is cashed out in vocation, in doing the specific work that is yours with the “good-humored inflexibility” to keep doing it when “the whole cry of voices is on the other side.” The sturdy New Hampshire lad Emerson praises — who “teams it, farms it, peddles, keeps a school, preaches, edits a newspaper… and always, like a cat, falls on his feet” — is worth “a hundred of these city dolls” precisely because he acts, tries all the professions, “does not postpone his life, but lives already.” Self-reliance is a theory of doing, and its proof is in the work.

8. Nothing external can save you — the erect posture

Now the doctrine turns and faces outward, at everything we lean on, and knocks the crutches away one by one. This is the principle that makes The Conduct of Life and Self-Reliance one continuous argument. “And so the reliance on Property, including the reliance on governments which protect it, is the want of self-reliance.” Note the reframe: leaning on property, reputation, patronage, party, even on travel, is not neutral — it is the absence of self-reliance, a measurable deficit. “Men have looked away from themselves and at things so long, that they have come to esteem the religious, learned, and civil institutions as guards of property… They measure their esteem of each other by what each has, and not by what each is.”

The most quoted formulation of this principle is the one that closes the essay, and it is a genuine philosophical claim, not a bromide: “Nothing can bring you peace but yourself. Nothing can bring you peace but the triumph of principles.” Read it against its context — a list of the “favorable events” that raise your spirits, “a political victory, a rise of rents, the recovery of your sick, or the return of your absent friend” — and its severity appears. Emerson is saying that to hang your equanimity on any external outcome is to have built your house on another man’s land. The image he gives for the alternative is physical: the man “who knows that power is inborn… throws himself unhesitatingly on his thought, instantly rights himself, stands in the erect position, commands his limbs, works miracles; just as a man who stands on his feet is stronger than a man who stands on his head.” Self-reliance is posture. It is standing on your own feet — not as a metaphor for stubbornness but as the literal geometry of strength.

9. Compensation — the ledger you cannot cheat

Underwriting the whole doctrine, and giving it a spine of iron that self-help entirely lacks, is Emerson’s law of Compensation — the subject of the essay that sits beside Self-Reliance in the First Series (and in this library in several forms). Its claim: the moral universe is a closed system of double-entry bookkeeping. “For everything you have missed, you have gained something else; and for everything you gain, you lose something.” There is no free lunch, no escaping the ledger, and — crucially — no need to try, because the ledger is just. “Every act rewards itself, or in other words integrates itself, in a twofold manner; first in the thing, or in real nature; and secondly in the circumstance, or in apparent nature.”

Why does this belong among the principles of self-confidence? Because it is what makes self-reliance safe to practise. If you believe the universe is arbitrary — that outcomes are handed out by luck, favour, or the crowd’s whim — then of course you will curry favour, hedge, conform, insure yourself against a capricious world. Compensation removes the ground for that anxiety. “So use all that is called Fortune. Most men gamble with her, and gain all, and lose all, as her wheel rolls. But do thou leave as unlawful these winnings, and deal with Cause and Effect, the chancellors of God. In the Will work and acquire, and thou hast chained the wheel of Chance.” The person who deals in cause and effect — who does the real work, tells the real truth — has “chained the wheel of Chance” and can afford to be indifferent to fortune, because the ledger will settle. This is the hidden hinge of the doctrine: self-reliance is rational only in a just universe, and Compensation is Emerson’s proof that the universe is just. Take away Compensation and self-reliance is reckless; keep it and self-reliance is simply the correct response to how reality actually works.

10. The self is portable — travel is a fool’s paradise

A smaller principle, but one of Emerson’s most quotable and most useful, aimed at the perennial fantasy of escape. “Traveling is a fool’s paradise. Our first journeys discover to us the indifference of places. At home I dream that at Naples, at Rome, I can be intoxicated with beauty, and lose my sadness. I pack my trunk, embrace my friends, embark on the sea, and at last wake up in Naples, and there beside me is the stern fact, the sad self, unrelenting, identical, that I fled from. My giant goes with me wherever I go.“

The principle generalises far past tourism: any attempt to fix an inner deficit by changing your outer circumstances is “a fool’s paradise.” The restlessness that sends us abroad is “a symptom of a deeper unsoundness affecting the whole intellectual action” — the same disease as imitation, the mind that “leans and follows the Past and the Distant” instead of standing where it is. “We imitate; and what is imitation but the traveling of the mind?” This is self-reliance as sufficiency of place: the conviction that the resources you need are wherever you are, because they are in you, and that the search for them elsewhere is a way of avoiding the only work that would actually help. It is the ancient Stoic insight — you cannot outrun yourself — pressed into service of the American doctrine of self-trust. And it has teeth for a mobile, optimising, grass-is-greener age that has turned the geographic fantasy into a career strategy and a dating app.

11. Society never advances — against the worship of the new

Emerson’s most bracing and least comfortable principle, and the one that most separates him from the progress-worship of his century and ours. “Society never advances. It recedes as fast on one side as it gains on the other… Society acquires new arts, and loses old instincts.” He drives it home with an image no technologist likes to hear: “The civilized man has built a coach, but has lost the use of his feet. He is supported on crutches, but lacks so much support of muscle. He has a fine Geneva watch, but he fails of the skill to tell the hour by the sun… His notebooks impair his memory; his libraries overload his wit.”

This is not reactionary grumbling; it is a strict application of Compensation (principle 9) to the domain of technology and civilisation. Every acquired capability is paid for with a lost one. The claim is conservation, not decline — “for everything that is given, something is taken” — and it lands as a permanent caution against confusing the improvement of tools with the improvement of persons: “the arts and inventions of each period are only its costume, and do not invigorate men… No greater men are now than ever were.” Galileo “with an opera-glass discovered a more splendid series of celestial phenomena than any one since”; Columbus “found the New World in an undecked boat.” The point for self-reliance is exact: do not mistake the growth of collective apparatus for the growth of your individual power. The apparatus is a wave — “the wave moves onward, but the water of which it is composed does not” — and to identify your worth with society’s accumulated arts is to lean, once again, on something outside yourself. For an age that treats each new technology as self-improvement by proxy, and never more so than in the age of intelligent machines, this is the most disquieting beam in the whole structure, and worth sitting with rather than dismissing.

12. The triumph of principles — self-trust as alignment, not assertion

The twelfth principle gathers the other eleven and states what self-reliance finally is, once you strip away the caricature. It is not the assertion of the ego against the world. It is alignment — the disciplined placing of the self in the current of a law it did not invent and cannot cheat. “Nothing can bring you peace but the triumph of principles.” Peace is not the reward of getting what you want; it is the by-product of standing in true relation to reality — to the Over-Soul (2), to Compensation (9), to the law of your own nature.

Emerson’s last word on the self-reliant person is therefore not a boast but a description of a burden freely taken: “And truly it demands something godlike in him who has cast off the common motives of humanity, and has ventured to trust himself for a taskmaster. High be his heart, faithful his will, clear his sight, that he may in good earnest be doctrine, society, law, to himself, that a simple purpose may be to him as strong as iron necessity is to others.” This is the antithesis of “you’ve got this.” It is closer to a monastic vow. To be “doctrine, society, law, to himself” is to have internalised the entire external apparatus of authority and to answer to it directly, without intermediaries — which is why the doctrine is so easily mistaken for arrogance and so rarely, in fact, practised. The self-reliant person has more law over him, not less; he has simply moved the courthouse inside. And the peace this buys is the only kind Emerson thinks is real: the peace of a life that has stopped negotiating with appearances and made itself “plastic and permeable to principles,” so that “by the law of nature” it “must overpower and ride all cities, nations, kings, rich men, poets, who are not.” Self-confidence, at the last, is not confidence in the self. It is confidence in the principles the self has consented to obey.

The cost, and the critics

Any honest account of this doctrine has to name what it costs, because Emerson names it and his critics have hammered it, and the criticism is in this library on purpose. Three charges recur, and each catches something real.

The charge of coldness. The most brilliant of Emerson’s critics, John Jay Chapman, whose 1898 essay sits in this library, admired him profoundly and still found at the centre of the work a chill — a philosophy so devoted to the solitary soul that it had little to say about love, grief, obligation, the thick human bonds that are not chosen. Emerson half-anticipates the charge and does not flinch from it: “I shun father and mother and wife and brother, when my genius calls me.” He tells the “foolish philanthropist” that “thy love afar is spite at home.” There is a hardness here that the mug-version launders away, and it is not an accident of tone; it follows from the doctrine. If nothing is sacred but the integrity of your own mind, then every relationship is provisional on that integrity, and Emerson accepts the consequence with a steadiness that can look like cruelty.

The charge of political naïveté — and worse. The library’s angle on reception and controversy holds the sharpest version. Emerson wrote, in Self-Reliance itself, a notorious jab at an abolitionist come “with his last news from Barbadoes” — a passage that reads very badly given that Emerson himself would become a committed antislavery voice within a decade. Critics have asked, reasonably, whether a philosophy that tells you to “grudge the dollar” you give to “such men as do not belong to me” can ground any politics of solidarity at all — whether self-reliance, taken neat, is not simply a genteel license for indifference to the weak. The Roosevelt Institute report in this library extends the worry into our own moment, tracing how “self-reliance” as a cultural ethos has been conscripted to justify the dismantling of collective provision — the doctrine weaponised into “you’re on your own.” That Emerson’s later career was a rebuke to the caricature (he did the opposite of retreat) does not dissolve the tension in the text. It is a real fault line, and the reader who papers over it has not read the essay.

The charge of incoherence. We have met it: Morley, More, and a long line of tidy minds who cannot forgive the contradictions. But this is the charge that self-destructs, because Emerson theorised his own inconsistency and made it doctrine (principle 4). You cannot convict a man of a crime he committed on purpose and explained in advance. The deeper reading — pursued in the library’s scholarship on Emersonian perfectionism, and by the philosopher Stanley Cavell, whose influence runs through several of the collected pieces — is that the contradictions are the method: truth as a series of circles, each provisional, none final. That is either a profound epistemology or an evasion of the duty to be pinned down, and honest readers can disagree. What no one can claim is that Emerson didn’t see it coming.

Why this outlived its century

The proof of a doctrine is what grows from it, and what grew from Self-Reliance is the intellectual history of American self-confidence — split, tellingly, into halves, each descendant keeping a different piece.

Nietzsche kept the self-overcoming and threw away the Over-Soul: he read Emerson devotedly (the library documents the influence in detail), lifted the becoming-self of principle 5 and the “insist on yourself” of principle 6, and hardened them into the will to power and the Übermensch — self-reliance with the metaphysics of connection stripped out and the metaphysics of struggle put in its place. The pragmatists — William James and John Dewey, both represented here by their own tributes — kept the empiricism of principle 7 (you cannot know your power till you try) and the anti-authoritarianism of the whole, and built from it a philosophy in which truth is what works, tested in action, owned by the individual knower. Dewey called Emerson “the Philosopher of Democracy” precisely because he had located authority in the ordinary individual soul. The psychology of self-efficacy — Albert Bandura’s foundational work sits in this library, alongside the resilience and self-belief research it spawned — kept the operational core: that belief in one’s own capability, built cumulatively through mastery experiences, is the best single predictor of what a person will attempt and achieve. That is principle 5 and principle 7, translated into the language of controlled experiment, shorn of the theology, and vindicated in the data.

And the self-help industry kept the slogan and threw away everything else — the epistemology, the Over-Soul, the ledger of Compensation, the demand that the self be built, the coldness, the cost. What it sells is the confidence without the discipline, the trust without the taskmaster, “trust thyself” with the iron string removed. That is why it doesn’t work, and why Emerson keeps having to be rediscovered underneath it. The mug is not Emerson. The mug is what’s left of Emerson after you remove everything that made him true.

What to do with it

If there is a single instruction to carry out of Self-Reliance into a networked, algorithmic, machine-augmented life, it is this: treat your own first-hand perception as admissible evidence, and treat the reflex of deference — to the crowd, to the feed, to the model, to your own past positions — as the thing to be interrogated, every time, before you act on it. Not to reject the secondhand: Emerson read voraciously, travelled, used the tools of his age. But to keep the line bright between the tool that extends your judgement and the habit that replaces it. The recommendation engine, the consensus of the timeline, the fluent answer from the machine — each is a “joint-stock company” offering the same ancient trade: security and belonging in exchange for the labour of thinking for yourself. Emerson’s essay is the price tag on that trade, written out in full, a hundred and eighty years before the machines got good enough to make the trade irresistible.

The self-confidence he taught is available to anyone, but it is not free and it is not a feeling. It is the slow accumulation of days on which you backed your own sight against the room and were right often enough to trust the faculty — the cumulative character of principle 5, cashed out in the vocation of principle 7, standing in the erect posture of principle 8, indifferent to fortune because of the ledger of principle 9, and grounded, if Emerson is right, in a source that is not yours at all. “Trust thyself: every heart vibrates to that iron string.” The string is iron because it is hard, and because it holds.

Source note. This essay is built on the ENSI Emerson research library — 94 downloaded primary and secondary documents across 15 angles. The primary quotations are from Self-Reliance, The Over-Soul, Compensation and Circles (Essays: First Series, 1841; Project Gutenberg and Internet Archive editions) and The Conduct of Life (1860). The critical and interpretive readings draw on the library’s scholarship angles — O’Dwyer and the Journal of Philosophy of Life*, the IJWOS and JETIR analyses of the doctrine, Kovalainen and the literature on Emersonian moral perfectionism, John Jay Chapman’s* Emerson and Other Essays*, Morley and Paul Elmer More on reception, Dewey’s and James’s tributes, the Emerson–Nietzsche scholarship, Bandura and the Frontiers/ERIC self-efficacy literature, and the Roosevelt Institute on the modern politics of self-reliance.

PART I — PRICING & ECONOMIC WARFARE

Moves that use money itself as a weapon — bleeding rivals, buying markets, and rewriting the unit economics of an entire category.

1. Predatory Pricing (The Scorched-Earth Discount)

Definition. The company deliberately prices a product below its own cost — not to make money, but to make it impossible for competitors to survive. Once rivals bleed out and exit, prices quietly rise again. This is the Flixbus-style move: flood a route with fares so low that every regional bus operator loses money trying to match them.

Mechanism. Predatory pricing works because the attacker has a deeper reserve of capital or cross-subsidy than the defender. It converts a war of products into a war of balance sheets, where the side that can absorb losses longest wins. The genius is asymmetry: the incumbent must defend every route, while the attacker chooses exactly where and when to strike, concentrating firepower until the local competitor’s cash runs dry.

Real-world example. Long-distance bus operator FlixBus expanded across Europe partly by undercutting national rail and legacy coach lines with fares as low as a few euros, sustaining losses on contested routes while building a dominant network. Amazon’s early willingness to sell books and later diapers (Quidsi/Diapers.com saga) at a loss to force acquisition is the canonical tech version.

Counter-move. Regulators sue for antitrust; rivals refuse to match on price and instead differentiate on service, or force the predator to keep bleeding by staying alive with cheaper capital.

2. Penetration Pricing (Buy the Market, Then Monetize)

Definition. Enter with prices so low that adoption explodes, capture a dominant share, and only later raise prices or introduce paid tiers once customers are hooked and switching feels painful.

Mechanism. Early low prices dissolve the friction of trying something new. The company is effectively buying market share with a discount and betting that the lifetime value of a locked-in customer dwarfs the margin sacrificed at acquisition. It works best where switching costs or habits accumulate over time, so today’s cheap customer becomes tomorrow’s captive one.

Real-world example. Uber and its rivals subsidized rides for years, offering fares below true cost to build rider habit and driver density before raising prices toward profitability. Streaming services routinely launch in new countries at a fraction of eventual pricing.

Counter-move. Competitors match the subsidy to deny share, or wait for the inevitable price hikes and poach the disillusioned customers.

3. The Freemium Land-Grab

Definition. Give the core product away free forever, monetize only a small fraction of power users, and use the enormous free base as both a marketing engine and a moat.

Mechanism. Free removes the single biggest barrier to adoption — the credit card. A massive free tier generates word-of-mouth, network effects, and data, while conversion of even 2–5% of users to paid can fund the whole operation. The free base is a wall: no competitor can charge for what you give away, so they must either match your generosity (burning cash) or attack a narrower niche.

Real-world example. Dropbox, Spotify, Slack, and Zoom all grew explosively by making the free tier genuinely useful, converting teams and heavy users into subscriptions once the tool became indispensable.

Counter-move. Rivals commoditize your free tier with an open-source or ad-supported clone, or out-premium you at the high end where the real money lives.

4. Cross-Subsidization (Rob Peter to Conquer Paul)

Definition. Use fat profits from one product, market, or customer segment to fund a below-cost assault in another, where you want to gain power.

Mechanism. A monopoly or high-margin cash cow becomes a war chest. The company can lose money indefinitely in a contested arena because the losses are invisibly financed elsewhere. This lets a firm enter a new market and behave as if gravity doesn’t apply, out-lasting focused specialists who have no second wallet.

Real-world example. Google funds dozens of free products (Maps, Docs, Android) with search-advertising profits, denying oxygen to standalone competitors who must charge for the same thing. Microsoft used Windows/Office cash to fund the browser wars.

Counter-move. Regulators unbundle the cash cow; focused rivals win the specific segment by being 10× better where the giant is merely “free and good enough.”

5. Razor-and-Blades Lock-In

Definition. Sell the durable device cheap (or at a loss) and make the real money on the recurring consumables, refills, or cartridges that only work with it.

Mechanism. The cheap razor lowers the entry barrier and pulls customers into an ecosystem; the proprietary blades create an annuity. Because the consumable is locked to the device, the company converts a one-time sale into a lifetime revenue stream and makes it expensive for the customer to defect after buying in.

Real-world example. Gillette (razors/blades), HP (printers/ink), Nespresso (machines/pods), and Keurig (brewers/K-cups) all subsidize hardware to sell high-margin refills forever.

Counter-move. Third parties sell compatible generic consumables; customers rebel against “ink tax” and switch to refillable or subscription-free alternatives.

6. Bundling (Strength in Numbers)

Definition. Package multiple products together at a price that makes buying the bundle irresistible versus buying any piece separately — using a strong product to carry weaker ones into the market.

Mechanism. Bundling leverages a must-have anchor product to force distribution of everything else in the box. It raises the perceived value, obscures the price of individual items, and denies rivals the oxygen to sell standalone competitors because the customer already “got it for free” in the bundle.

Real-world example. Microsoft Office bundled Word, Excel, and PowerPoint, crushing standalone WordPerfect and Lotus. Amazon Prime bundles shipping, video, music, and photos so the whole is unassailable by any single-purpose rival.

Counter-move. Regulators force unbundling; nimble competitors win by being dramatically better at the one component customers care about most.

7. Premium Anchoring (The Price Umbrella)

Definition. Position deliberately at the very top of the price range to define the category’s ceiling, make everything else look cheap, and capture the highest-margin customers and brand halo.

Mechanism. A high price is a signal of quality and status; it anchors customer expectations and lets the leader harvest enormous margins while competitors fight in the discounted mud below. The premium tier also funds R&D and marketing that reinforce the very superiority the price implies — a self-fulfilling loop.

Real-world example. Apple’s premium pricing captures the vast majority of smartphone industry profits despite a minority of unit share. Tesla, Rolex, and Louis Vuitton all use price itself as the primary signal of desirability.

Counter-move. “Good enough” disruptors attack from below with 80% of the value at 40% of the price, slowly eroding the premium’s justification.

8. Dynamic & Personalized Pricing

Definition. Charge each customer, moment, or context a different price calculated in real time to extract the maximum each is willing to pay.

Mechanism. By harvesting data on demand, scarcity, and individual behavior, the firm captures “consumer surplus” that flat pricing leaves on the table. It maximizes revenue per transaction and lets the company subsidize price-sensitive segments while gouging the desperate or the loyal — all invisibly.

Real-world example. Airlines, Uber’s surge pricing, Amazon’s fluctuating listings, and hotel revenue-management systems all reprice constantly based on demand signals.

Counter-move. Price-comparison tools and public backlash force transparency; competitors win goodwill by promising simple, fair, fixed prices.

PART II — TECHNOLOGY & PRODUCT SUPREMACY

Moves that win by building something rivals simply cannot match — or by owning the rules everyone else must build on.

9. The Technology Moat (Be Genuinely Best)

Definition. Invest so heavily and so early in a hard technical capability that you produce something meaningfully better than anyone else can, and keep the lead by out-investing the field.

Mechanism. Deep technical superiority creates a quality gap customers can feel, commanding premium prices and loyalty. The moat compounds: revenue funds more R&D, which widens the lead, which funds more R&D. Where the technology is genuinely hard (rockets, chips, models), the capital and talent required become a barrier that money alone can’t quickly cross.

Real-world example. SpaceX’s reusable rockets, TSMC’s leading-edge chip fabrication, ASML’s EUV lithography, and NVIDIA’s GPU/CUDA stack each represent a technical lead measured in years that rivals cannot simply buy their way past.

Counter-move. Rivals leapfrog with a new paradigm that makes your hard-won lead irrelevant, or commoditize the capability through open collaboration.

10. Owning the Standard

Definition. Get your proprietary format, protocol, or interface adopted as the industry standard so that everyone else must build on your foundation — and pay you rent to do so.

Mechanism. Standards create winner-take-all lock-in: once the ecosystem coalesces around one way of doing things, the cost of switching is collective and therefore nearly infinite. The standard-owner sits at the toll booth, extracting licensing fees, steering the roadmap, and ensuring compatibility flows through them.

Real-world example. Qualcomm’s cellular patents, Dolby audio, the MP3 licensors, and Adobe’s PDF (before it opened) all turned a format into a perpetual tax on an entire industry.

Counter-move. A rival coalition backs an open, royalty-free standard (e.g., open codecs) to strand the proprietary toll booth.

11. The Patent Thicket

Definition. Blanket a technology area with a dense web of overlapping patents so that no competitor can build in the space without infringing something you own.

Mechanism. Individual patents can be designed around; a thicket cannot. By owning hundreds of interlocking claims, the firm creates a legal minefield that raises rivals’ costs, delays them in court, and forces them into licensing deals or cross-licensing on your terms. It weaponizes the legal system as a barrier to entry.

Real-world example. The smartphone patent wars saw Apple, Samsung, and others amass and litigate vast portfolios; Qualcomm and IBM built thickets that generate billions in licensing revenue.

Counter-move. Rivals form patent pools, buy defensive portfolios, or lobby for patent reform; open-source patent pledges neutralize the threat.

12. Vertical Integration (Own the Whole Stack)

Definition. Control every layer of your value chain — from raw materials to end customer — so no supplier or distributor can hold you hostage, and so you capture margin at every step.

Mechanism. Owning the stack removes dependency and its associated risk, guarantees supply, and lets the firm optimize across layers in ways fragmented competitors cannot. It also raises the barrier to entry, because a new rival must replicate not one business but the entire chain.

Real-world example. Tesla builds its own batteries, chips, software, and retail stores; Apple designs its own silicon, OS, and stores; Amazon owns warehousing, logistics, and cloud. Standard Oil pioneered this by controlling pipelines and refining alike.

Counter-move. Focused specialists at each layer out-innovate the generalist; the integrator’s rigidity becomes a liability when one layer shifts fast.

13. The Platform Play (Become the Ground Others Stand On)

Definition. Transform your product into a platform that third parties build on top of, so their success becomes your success and their investment becomes your lock-in.

Mechanism. A platform harnesses the labor and creativity of thousands of outside developers, multiplying the value of the core product without proportional cost. Every app built on the platform deepens the moat, because that ecosystem cannot easily be replicated or moved. The platform owner sets the rules and takes a cut of the economy it hosts.

Real-world example. Apple’s App Store, Microsoft Windows, Salesforce’s AppExchange, and Shopify’s app ecosystem all turned products into economies the owner taxes and governs.

Counter-move. Developers revolt over high “taxes” and migrate to open platforms; regulators force lower fees and sideloading.

14. First-Mover Land Rush

Definition. Move first and fast into a new market to claim the best customers, mindshare, and resources before anyone realizes the opportunity exists.

Mechanism. Being first lets you define the category, set customer expectations in your image, and lock up scarce assets (spectrum, real estate, key partnerships, top talent). Early leadership can compound into network effects and brand default before competitors even arrive.

Real-world example. Amazon in e-commerce, Google in search, and Coinbase in mainstream US crypto each seized durable advantages by being early and aggressive.

Counter-move. Fast-followers learn from the pioneer’s expensive mistakes and win with a refined second version (the “second-mouse-gets-the-cheese” gambit).

15. Deliberate Obsolescence & The Upgrade Treadmill

Definition. Design products with a limited useful life or a relentless upgrade cadence so customers must keep buying to stay current.

Mechanism. By tying performance, compatibility, or support to the newest version, the firm converts durable goods into recurring purchases. Ecosystem effects (new software that needs new hardware) accelerate the treadmill, keeping the revenue flywheel spinning.

Real-world example. Annual smartphone releases, fashion’s seasonal cycles, and software that drops support for older hardware all keep customers upgrading.

Counter-move. Right-to-repair laws, durable-goods challengers, and second-hand markets slow the treadmill and shame the practice.

PART III — DISTRIBUTION & CHANNEL CONTROL

Whoever owns the road to the customer owns the customer. These moves seize the channel.

16. Exclusive Distribution Lock-Up

Definition. Sign deals that make you the only option in a given channel, shelf, or venue — legally excluding rivals from reaching customers there.

Mechanism. Even a superior competitor is powerless if it cannot reach the buyer. By locking channels — with volume rebates, exclusivity clauses, or category-captain status — the firm starves rivals of distribution, the one thing money can’t quickly manufacture.

Real-world example. Beverage giants’ exclusive pouring rights at restaurants and stadiums, and historic cases where dominant firms tied retailer rebates to not stocking rivals, illustrate channel exclusion.

Counter-move. Rivals open new channels (direct-to-consumer, e-commerce) that bypass the locked ones entirely.

17. Owning the Default Position

Definition. Pay or engineer your way into being the pre-set, out-of-the-box choice, because the overwhelming majority of users never change defaults.

Mechanism. Defaults exploit human inertia. Being the default is worth more than being the best, because it captures the vast passive majority automatically. It converts a distribution deal into near-monopoly usage with zero ongoing persuasion.

Real-world example. Google pays Apple tens of billions annually to be the default Safari search engine; pre-installed apps on phones dominate their categories through sheer placement.

Counter-move. Regulators mandate “choice screens” that force users to pick, breaking the default’s grip.

18. Shelf-Space Saturation

Definition. Flood every available slot — physical shelf, app store category, ad inventory — with so many of your own SKUs that competitors literally cannot find room.

Mechanism. Finite shelf space is zero-sum. By occupying most of it with a portfolio of brands (often disguised as competitors of each other), the firm crowds out genuine rivals and controls what the customer even sees. Choice is an illusion when every option is yours.

Real-world example. Consumer-goods giants like P&G and Unilever, and cereal makers, historically saturated shelves with dozens of brands owned by the same parent.

Counter-move. E-commerce’s “infinite shelf” and direct-to-consumer brands escape the physical bottleneck entirely.

19. Owning the Last Mile

Definition. Control the final, hardest, most expensive link to the customer — delivery, installation, the physical storefront — so that others must route through you.

Mechanism. The last mile is the costliest and least replicable part of many value chains. Whoever owns it controls the customer relationship, the data, and the timing, and can charge everyone upstream for access. It is a moat built from logistics and concrete, not code.

Real-world example. Amazon’s logistics network, telecom “last-mile” cables into homes, and utility grids all confer control over everything that must reach the end user.

Counter-move. Aggregators and gig networks assemble a virtual last mile; rivals partner to pool delivery and share the cost.

PART IV — MERGERS, ACQUISITIONS & CAPITAL AS A WEAPON

Moves that reshape the board by buying pieces off it — or by simply having more chips than everyone else.

20. The Roll-Up (Consolidate a Fragmented Market)

Definition. Systematically buy up dozens of small players in a fragmented industry, merge them into one dominant entity, and reap scale, pricing power, and efficiency.

Mechanism. Fragmented markets have no price discipline and duplicated overhead. A roll-up replaces many weak hands with one strong one, gaining purchasing leverage, eliminating redundant costs, and gaining pricing power over customers and suppliers who now face fewer alternatives.

Real-world example. Private-equity roll-ups of veterinary clinics, dental practices, funeral homes, and HVAC companies; Waste Management in trash hauling; countless “buy-and-build” strategies.

Counter-move. Antitrust review of serial acquisitions; new independents spring up to serve customers alienated by the consolidator.

21. The Killer Acquisition

Definition. Buy a promising young competitor specifically to shut it down or absorb it before it can threaten your core business.

Mechanism. It is cheaper to buy a future threat for millions than to fight it for billions later. By acquiring nascent rivals early, the incumbent removes tomorrow’s disruptor while it is still affordable, and folds its talent or technology (or simply kills it) to protect the mothership.

Real-world example. Facebook’s acquisitions of Instagram and WhatsApp are the textbook cases regulators cite; pharma firms buying and shelving competing drug pipelines is another.

Counter-move. Antitrust scrutiny of “nascent competitor” deals; founders who refuse to sell and out-execute the incumbent.

22. The Acqui-Hire Talent Raid

Definition. Acquire a small company primarily to absorb its people, gutting a competitor’s or a promising startup’s talent base in one stroke.

Mechanism. In knowledge industries, the team is the asset. Buying the whole company is a fast, clean way to hire a proven unit and simultaneously deny that talent to everyone else. It removes a potential competitor and strengthens you in a single transaction.

Real-world example. Big Tech’s frequent purchases of tiny AI and hardware startups largely for their engineering teams; the recent wave of AI “reverse acqui-hires” of founding teams.

Counter-move. Non-competes’ erosion and equity retention packages keep talent independent; rivals counter-recruit the same teams.

23. Capital as a Weapon (Outspend to Outlast)

Definition. Raise or deploy so much capital that you can simply outspend every rival on growth, marketing, and subsidies until they run out of money.

Mechanism. With a bottomless war chest, a firm can turn any market into a capital-endurance contest it is designed to win. It blitzes on advertising, undercuts on price, and floods on hiring, forcing rivals to either match the burn (and die) or retreat. Money becomes the moat.

Real-world example. SoftBank’s Vision Fund pouring billions into WeWork, Uber, and others; the “blitzscaling” of well-funded startups that spent rivals into oblivion.

Counter-move. Capital discipline reverses (rates rise, funding dries up) and the over-funded giant collapses under its own burn; lean rivals survive on real economics.

24. Backward Integration (Buy Your Suppliers)

Definition. Acquire your key suppliers so you control your inputs, deny them to competitors, and capture their margin.

Mechanism. Owning the supply chain guarantees your own access to a scarce or critical input while potentially cutting off rivals who depended on the same source. It converts a cost center into a controlled asset and a competitive weapon.

Real-world example. Tesla’s investments in lithium and battery supply; automakers buying chip capacity; Amazon building its own delivery fleet to escape UPS/FedEx dependency.

Counter-move. Rivals secure alternative suppliers or vertically integrate themselves; suppliers refuse exclusivity to preserve their broader market.

PART V — BRAND, PERCEPTION & PSYCHOLOGICAL WARFARE

Markets are won in the mind before they are won in the wallet. These moves manipulate belief.

25. Category Creation (Own a Word in the Mind)

Definition. Invent and name a new category, then position yourself as its definitional leader, so that the category and your brand become synonymous.

Mechanism. The company that names the category owns it. By defining the problem and the solution, the firm frames all competitors as imitators of itself. Owning a word in the customer’s mind (”the CRM,” “the search engine”) is the most durable moat of all because it lives in language.

Real-world example. Salesforce (”cloud CRM”), Red Bull (”energy drink”), Xerox and Google becoming verbs, HubSpot (”inbound marketing”).

Counter-move. Rivals reframe the category on new terms that make the incumbent’s definition sound dated.

26. Manufactured Scarcity & Exclusivity

Definition. Deliberately limit supply or access to inflate desire, status, and price — making the product a symbol precisely because not everyone can have it.

Mechanism. Scarcity triggers loss aversion and status-seeking. By capping availability (limited drops, waitlists, invite-only), the firm converts a functional product into a positional good whose value derives from exclusion. Demand outruns supply, and the brand accrues mystique.

Real-world example. Hermès Birkin bags, Supreme drops, Ferrari’s production caps, and invite-only product launches (early Gmail, Clubhouse).

Counter-move. Abundant “democratized” alternatives satisfy the underserved demand and reframe scarcity as gatekeeping.

27. Thought Leadership & Mindshare Capture

Definition. Dominate the conversation in your field through content, research, conferences, and evangelism so that your worldview becomes the industry’s default frame.

Mechanism. Whoever shapes how an industry thinks shapes what it buys. By setting the agenda, defining best practices, and educating the market, the firm makes its own product the natural conclusion of every discussion. Mindshare converts to market share downstream.

Real-world example. McKinsey’s frameworks, a16z’s and Stripe’s publishing, NVIDIA’s GTC conference defining the AI narrative, Gartner’s Magic Quadrant shaping enterprise buying.

Counter-move. Contrarian challengers win attention by attacking the reigning orthodoxy the leader established.

28. FUD (Fear, Uncertainty, and Doubt)

Definition. Seed doubt about a competitor’s reliability, safety, longevity, or roadmap to scare risk-averse customers back to the “safe” incumbent choice.

Mechanism. In high-stakes purchases, fear beats features. By amplifying uncertainty (”nobody got fired for buying us,” “will that startup even exist next year?”), the incumbent exploits buyers’ risk aversion to freeze them in place, denying rivals the benefit of the doubt they need to win trials.

Real-world example. The classic “IBM/Microsoft” FUD playbook against open-source and smaller vendors; enterprise incumbents warning of the risks of switching to a newer challenger.

Counter-move. Challengers publish proof — uptime data, big-name references, guarantees — that turns the fear around on the aging incumbent.

29. Vaporware & Preannouncement

Definition. Announce a product long before it exists (or when it never will) to freeze customers who might otherwise buy a competitor’s shipping product.

Mechanism. A credible promise of “something better coming soon” makes buyers wait. By preannouncing, the incumbent chills the market for a rival’s real product, buying time to actually build a response or simply to protect current sales. It weaponizes the future against the present.

Real-world example. Historic tech “vaporware” announcements timed to blunt competitors’ launches; the term itself was coined amid 1980s software rivalries.

Counter-move. Rivals ship, ship again, and publicly track the incumbent’s broken promises to erode credibility.

30. Astroturfing & Manufactured Consensus

Definition. Create the appearance of grassroots enthusiasm, reviews, or public support that is actually orchestrated and funded by the company.

Mechanism. Social proof drives behavior; people follow perceived crowds. By manufacturing the appearance of a movement — fake reviews, paid advocates, seeded “independent” voices — the firm bends perception of what’s popular, safe, or true, steering the herd toward itself or against a rival.

Real-world example. Documented fake-review economies on marketplaces, and lobbying front groups posing as citizen movements in policy fights.

Counter-move. Platform crackdowns, verified-purchase reviews, and journalistic exposure that turns the deception into a scandal.

PART VI — TALENT & KNOWLEDGE MONOPOLIES

In knowledge economies, cornering the people is cornering the market.

31. The Talent Monopoly

Definition. Hire so aggressively in a scarce specialty that you corner the available expertise, starving competitors of the people they need to compete.

Mechanism. When a capability depends on a few hundred experts worldwide, whoever employs most of them owns the capability. Lavish compensation concentrates rare talent, which accelerates your progress and simultaneously handicaps everyone else who can’t staff the same effort.

Real-world example. The bidding wars for top AI researchers, where a handful of labs pay eight-figure packages to lock up the field’s leading minds; quant funds cornering specialized math talent.

Counter-move. Rivals grow talent internally, tap overlooked geographies, or use tooling that reduces the number of experts required.

32. Poaching the Key Person

Definition. Recruit a competitor’s irreplaceable individual — the star engineer, the rainmaking salesperson, the visionary designer — to cripple them and empower yourself.

Mechanism. Some organizations have a keystone: remove that one person and the arch weakens. Poaching the keystone can transfer knowledge, relationships, and momentum in a single hire, delivering a double blow — you gain exactly what your rival loses.

Real-world example. High-profile executive and engineering defections in tech and finance that shifted entire product roadmaps and client books between firms.

Counter-move. Retention equity, non-solicits, and building institutional (not individual) knowledge so no single departure is fatal.

33. Non-Competes & Knowledge Fencing

Definition. Use legal contracts — non-competes, NDAs, IP assignment, trade-secret suits — to fence in your people and knowledge so they can’t flow to rivals.

Mechanism. By legally restricting where employees can go and what they can share, the firm turns its accumulated know-how into a defensible asset. Even the threat of litigation deters competitors from hiring your people or copying your methods, freezing the knowledge inside your walls.

Real-world example. Trade-secret and non-compete litigation across tech and manufacturing; the ongoing regulatory battles over non-compete enforceability in the US.

Counter-move. Courts and regulators void overbroad non-competes; talent migrates to jurisdictions (like California) that refuse to enforce them.

PART VII — REGULATORY & LEGAL MANEUVERS

The rules of the game are themselves a battlefield. These moves capture the referee.

34. Regulatory Capture

Definition. Influence the regulators and rule-makers so that the rules end up protecting your position rather than constraining it.

Mechanism. Incumbents fund lobbying, place alumni in agencies, and shape technical standards so that regulation — nominally for public good — quietly entrenches them. Complex rules favor those with the lawyers and compliance budgets to handle them, turning oversight into a moat against smaller challengers.

Real-world example. Heavily lobbied industries (finance, telecom, pharma, defense) where established players shape the very regulations that govern them; occupational licensing that limits new entrants.

Counter-move. Public-interest advocacy, transparency mandates, and challenger coalitions that lobby for pro-competition rules.

35. Lawfare (Litigation as a Weapon)

Definition. Use lawsuits — meritorious or not — to drain a rival’s cash, time, and attention, regardless of who ultimately wins.

Mechanism. Litigation is asymmetric when one side has vastly more resources: the process itself is the punishment. A well-funded firm can bury a smaller rival in discovery and legal fees, freezing their fundraising and roadmap until they settle, fold, or bleed out — the verdict almost beside the point.

Real-world example. Patent-troll suits against startups; deep-pocketed incumbents filing serial suits against disruptive entrants to slow them down.

Counter-move. Anti-SLAPP laws, fee-shifting, litigation-finance backing for the underdog, and public sympathy for the “David.”

36. The Compliance Moat

Definition. Turn burdensome regulation into your competitive advantage by mastering it so thoroughly that compliance itself becomes a barrier rivals can’t clear.

Mechanism. When entering a market requires expensive licenses, audits, certifications, or capital reserves, the firm that has already paid those costs enjoys a moat measured in years and millions. New entrants face a wall of compliance before they can even start, so the incumbent welcomes — even lobbies for — more regulation.

Real-world example. Banking charters, medical-device FDA approval, defense security clearances, and fintechs that turned money-transmitter licensing into a barrier against copycats.

Counter-move. “Regulatory arbitrage” challengers operate in lighter-touch jurisdictions or novel legal categories the rules haven’t caught up to.

37. Setting the Rules You Can Meet (Standards Gaming)

Definition. Lobby for standards, certifications, or thresholds calibrated to exactly what you can do and your rivals cannot — dressing a competitive attack as public interest.

Mechanism. By shaping the specifics of a mandate (an emissions limit, a safety spec, a data rule), the firm ensures the requirement is trivial for itself and crippling for competitors. The rule looks neutral and virtuous while functioning as a targeted weapon.

Real-world example. Emissions and efficiency standards shaped to favor certain technologies; safety or data-privacy rules whose compliance cost only large incumbents can bear.

Counter-move. Rivals expose the self-serving design and lobby for outcome-based rather than prescriptive standards.

PART VIII — DATA, LOCK-IN & SWITCHING COSTS

Make leaving so painful that customers stay even when they’d rather go.

38. The Data Moat (Network Effects of Information)

Definition. Accumulate proprietary data at a scale rivals can’t match, so your product gets better with use in ways competitors can never catch up to.

Mechanism. More users generate more data, which improves the product, which attracts more users — a compounding loop. Because the data is proprietary and accumulates over time, a late entrant faces not just a better product but an insurmountable history of learning baked into it.

Real-world example. Google’s search relevance improving from query data, Waze’s traffic data, credit bureaus, and recommendation engines that improve with every interaction.

Counter-move. Data-portability regulation, synthetic data, and privacy-preserving techniques that let challengers bootstrap without the incumbent’s history.

39. Switching-Cost Engineering

Definition. Deliberately architect the product so that leaving means losing data, retraining staff, rebuilding integrations, or breaking workflows — making defection prohibitively painful.

Mechanism. Customers stay not because they love you but because leaving costs too much. By embedding your product deep into the customer’s operations — data formats, integrations, learned habits, contractual terms — the firm raises the exit price far above any rival’s advantage, locking in revenue even as satisfaction wanes.

Real-world example. Enterprise software (ERP systems like SAP/Oracle) where migration takes years; cloud egress fees; proprietary file formats that trap documents.

Counter-move. Open standards, migration tooling, and interoperability mandates that lower the exit cost.

40. The Walled Garden

Definition. Build a closed ecosystem where hardware, software, and services only work together, so buying into one part pulls you into all of it and makes leaving mean abandoning everything.

Mechanism. Each product in the garden increases the switching cost of the whole. The more of the ecosystem a customer adopts, the more expensive and disruptive it becomes to leave any single piece, because they’re all interlocked. Lock-in is achieved not by one anchor but by a web of them.

Real-world example. Apple’s ecosystem (iPhone, Mac, Watch, AirPods, iMessage, iCloud) where each device works best with the others and defection means replacing them all; console gaming ecosystems.

Counter-move. Interoperability regulation (forcing messaging or charging standards); rivals offering “we’ll pay your switching cost” migration deals.

41. Integration Entrenchment (Become Load-Bearing)

Definition. Wire your product so deeply into the customer’s other systems and workflows that removing it would break dozens of things — making you infrastructure, not a vendor.

Mechanism. A tool that many other tools depend on becomes structurally impossible to remove without collateral damage. By encouraging integrations, APIs, and dependencies, the firm makes itself load-bearing: ripping it out risks toppling everything built on top, so customers keep paying rather than risk the collapse.

Real-world example. Payment processors, identity providers (Okta/Auth0), and databases that dozens of downstream systems depend on; Slack/Teams woven into hundreds of workflow integrations.

Counter-move. Abstraction layers and middleware that let customers swap the underlying provider without breaking the dependencies.

PART IX — SUPPLY, RESOURCE & INFRASTRUCTURE CONTROL

Own the scarce inputs everyone needs and you tax the entire industry.

42. Cornering the Supply (Resource Capture)

Definition. Buy up or lock in control of a scarce, critical raw material or input so that rivals must come to you — or go without.

Mechanism. If everyone needs an input and you control its supply, you control everyone. By securing mines, reserves, capacity, or long-term supply contracts, the firm can set prices, prioritize its own needs, and choke competitors at the source, independent of how good their products are.

Real-world example. OPEC’s oil coordination, China’s rare-earth dominance, De Beers’ historic diamond control, and firms locking up long-term chip-fab or battery-mineral capacity.

Counter-move. Substitution, recycling, new supply discoveries, and strategic reserves that break the chokehold over time.

43. Capacity Pre-Emption

Definition. Book or build so much production capacity ahead of demand that competitors are locked out of the ability to scale even if their product is superior.

Mechanism. In capacity-constrained industries, whoever reserves the scarce manufacturing slots wins regardless of design quality. By pre-buying years of a supplier’s output, the firm denies rivals the physical means to produce at scale, converting a supply advantage into a market advantage.

Real-world example. Apple pre-buying leading-edge chip capacity from TSMC; hyperscalers reserving years of GPU and data-center power capacity ahead of AI demand.

Counter-move. Rivals fund new capacity, sign their own long-term deals, or design around the constrained component.

44. Owning the Infrastructure Everyone Rents

Definition. Own the underlying infrastructure — cloud, rails, pipes, grid, marketplace — that your own competitors must use, so you profit even when they win.

Mechanism. The “toll road” position means you earn a cut of all traffic, including your rivals’. You gain visibility into their operations, can prioritize your own services, and collect rent from the entire market. Competing at the application layer becomes secondary when you own the layer beneath it.

Real-world example. Amazon AWS hosting companies that compete with Amazon’s retail; app stores taxing apps that rival the platform’s own; exchanges and payment rails that everyone must transact through.

Counter-move. Multi-homing across providers, open infrastructure alternatives, and antitrust “structural separation” that forces the owner out of competing businesses.

PART X — ECOSYSTEM, NETWORK & FLYWHEEL EFFECTS

The most durable power is self-reinforcing power — moves that make winning cause more winning.

45. The Network-Effect Flywheel

Definition. Build a product whose value to each user increases as more users join, so growth becomes self-reinforcing and the leader becomes nearly unassailable.

Mechanism. Network effects create winner-take-most dynamics: each new user makes the product more valuable, attracting more users, in a loop that starves smaller rivals of the critical mass they need. Past a tipping point, the leader’s dominance compounds automatically — the market tips to a single winner.

Real-world example. Telephone and fax networks historically; today, social networks (Facebook), marketplaces (eBay), payment networks (Visa), and communication tools whose value is the network itself.

Counter-move. Niche networks that serve an underserved segment better, interoperability that lets small networks combine, and multi-homing that dilutes any single network’s lock.

46. Two-Sided Marketplace Domination

Definition. Own the marketplace that connects two groups (buyers/sellers, riders/drivers) so densely that neither side can afford to be anywhere else.

Mechanism. Liquidity begets liquidity: buyers go where the sellers are and sellers go where the buyers are, so the marketplace with the most of both becomes the only viable venue. The operator sits in the middle, taking a fee on every transaction and controlling the rules, ranking, and data of the entire economy.

Real-world example. Amazon Marketplace, Uber, Airbnb, App Store, and stock exchanges — each a two-sided market whose density makes alternatives feel empty.

Counter-move. Sellers organizing to multi-home or go direct-to-consumer; challengers subsidizing one side to bootstrap liquidity elsewhere.

47. Commoditize Your Complement

Definition. Drive the price of whatever is sold alongside your product toward zero — often by funding free or open-source alternatives — so that more spending flows to you.

Mechanism. Demand for your product rises when its complements get cheaper. By deliberately commoditizing the complementary layer (giving it away, backing open-source), the firm expands its own market and denies any rival the chance to build power in that adjacent space. You make the thing next to you free so the thing you sell becomes more valuable.

Real-world example. Google backing free Android to commoditize mobile hardware and keep search dominant; IBM and others funding Linux to commoditize the OS beneath their services; Meta open-sourcing AI models to commoditize the model layer.

Counter-move. The “commoditized” layer finds independent monetization, or a rival commoditizes your layer in return.

48. The Developer/Partner Ecosystem Lock

Definition. Cultivate a huge ecosystem of developers, integrators, and partners whose businesses and skills are invested in your platform, so their livelihoods defend your moat.

Mechanism. When thousands of people have built careers, certifications, and companies on your platform, they become an army of advocates and a switching cost in human form. Migrating away means abandoning an entire community’s accumulated expertise and integrations — so the ecosystem itself resists change.

Real-world example. Salesforce’s consultant and ISV ecosystem, AWS’s certified-architect economy, SAP’s implementation partners, and gaming platforms’ modding communities.

Counter-move. Rivals court the same partners with better economics; open ecosystems let partners hedge across multiple platforms.

49. The Cult Brand & Community Moat

Definition. Cultivate such deep identity, belonging, and emotional loyalty that customers defend the brand, evangelize it for free, and refuse alternatives on principle.

Mechanism. When a brand becomes part of a customer’s identity, the relationship transcends features and price. The community self-polices, recruits new members, and forgives missteps — creating a moat made of belief that no rival can buy with a better spec sheet. Loyalty becomes irrational, and therefore durable.

Real-world example. Apple, Harley-Davidson, Tesla, CrossFit, and Lego’s adult-fan community — customers who tattoo the logo and argue for the brand unpaid.

Counter-move. Brands go stale or betray the community’s values, opening the door for an authentic challenger to capture the disillusioned faithful.

50. The Aggregation Play (Own the Demand)

Definition. Position yourself between customers and a fragmented supply base, own the customer relationship and demand, and commoditize the suppliers who must come through you to reach anyone.

Mechanism. By controlling demand — the customers’ attention and default choice — the aggregator flips power over a fragmented supply side. Suppliers compete to be listed, driving their own margins down, while the aggregator takes a cut and owns the data and the relationship. Whoever owns demand dictates terms to everyone who wants to reach it.

Real-world example. Google and Meta aggregating advertising demand over publishers; Booking.com and Expedia over hotels; food-delivery apps over restaurants; Netflix over content.

Counter-move. Suppliers build direct relationships and brands to escape the aggregator; regulators challenge self-preferencing and fee extraction.

HOW TO READ THIS ENCYCLOPEDIA

These fifty moves are not fifty separate tricks — they are variations on a handful of deep principles of power:

• Asymmetry — fight where you’re strong and they’re weak (predatory pricing, cross-subsidy, lawfare).

• Lock-in — make leaving cost more than staying (switching costs, walled gardens, razor-and-blades).

• Compounding — arrange things so winning causes more winning (network effects, data moats, flywheels).

• Control of the chokepoint — own the one thing everyone must pass through (standards, infrastructure, last mile, supply).

• Perception — win the mind before the market (category creation, FUD, mindshare, cult brand).

The most powerful companies rarely rely on one move. They stack them — Apple runs premium anchoring and the walled garden and the platform play and the cult brand and capacity pre-emption at once — so that even if one moat is breached, four more remain.

The human world runs on a force exactly like this, and we have spent centuries calling it by a dozen different names. We call it “attention,” “trust,” “credibility,” “brand,” “reputation,” “narrative,” “mindshare,” “network effects,” “social capital.” These are not separate things. They are the same force seen from different angles, and the quantity underneath all of them is consciousness — the finite, directed capacity of minds to attend, to believe, and to bestow support. Consciousness is the gravity of the human world. It is pathetically weak in any single instance — anyone can ignore you, doubt you, walk away from you — and it is the only force that decides which products, ideas, companies, movements, currencies and people accumulate mass and which dissolve back into noise.

Every builder learns this the hard way, usually too late. You can rebuild Figma. The software is not magic; a competent team can reproduce the features, the rendering engine, the interface — you can clone the whole thing in a quarter. What you cannot rebuild is the fact that the world’s designers already look there. That attention, credibility and trust have already pooled around it, bending the decisions of every new designer toward the place the mass already sits. You can build a better product and still be finished, because if nobody knows you exist, if you do not have the attention, if you do not have the credibility, then it does not matter how good the thing is. The product is matter. The attention is gravity. And matter without gravity just drifts out into the dark, unseen, unpicked, unremembered.

When people say “some people simply attract support,” they are not being poetic. They are describing a physical intuition with the wrong vocabulary. Some people, some products, some ideas carry more gravitational mass in the medium of consciousness, and mass — as Newton and Einstein both understood — attracts more mass. By giving your attention to something you are giving it support; you are transferring a piece of your finite consciousness and adding to that thing’s pull. That is why attention is not a nice-to-have layered on top of the real work. It is the real work. It is the mass term in the only equation that decides what survives.

This is not a metaphor dressed up as a law. It is a law wearing the costume of a metaphor. The claim of this essay is that the parallel between gravity and consciousness is exact, feature by feature, and that once you see it you can stop guessing at why some things win and start engineering the field so it bends toward you. Below are sixteen principles. Each one takes a real, specific property of gravity and shows its precise counterpart in the physics of the human world — and each is grounded in the actual literature: general relativity and emergent gravity, the network science of preferential attachment, the economics of attention, trust and superstars, the sociology of collective consciousness, and the hard experiments on how success really propagates.

Before the sixteen, three features of gravity are worth naming, because the analogy is strong exactly where they line up. First, gravity couples to mass — bigger bodies pull harder. Second, it is one-signed — always attractive, never repulsive, with no negative mass to push things apart — which is why it never cancels and therefore concentrates without limit. Third, in Einstein’s picture it is not a push at all but the shape of space itself: mass tells spacetime how to curve, and the curvature tells everything else how to move.

Hold onto that third one, because it is the deepest idea here. A planet orbiting a star is not tugged on a string. It travels the straightest possible path through a valley the star’s mass has carved into the geometry. A body heavy enough does not need to push you — it changes the terrain so that moving toward it feels like going straight. That is what a dominant brand, a trusted name, a famous idea does to everyone around it. And there is even a frontier hint worth pocketing: Erik Verlinde’s work on emergent, entropic gravity argues that gravity may not be fundamental at all, but a statistical property of information — the way temperature emerges from moving molecules. Whether or not it holds in physics, it describes social gravity perfectly. A force that is utterly real and utterly universal, and yet nothing more than what happens when enough information is arranged a certain way.

The argument in brief — the sixteen laws of social gravity:

1. Attention is mass. Nothing has consequence until it first has attention; it is the one scarce resource everything else is smelted from.

2. The weakest force builds the world. Trivially overridden up close, unstoppable at scale — because, like gravity, it never cancels.

3. Attention only ever attracts. It has one sign — even hatred feeds it — so it can never spread evenly, only concentrate into a power law.

4. Mass attracts mass. The rich get richer by preferential attachment; early leads accrete into near-permanent dominance.

5. Consciousness curves the space others move through. A massive brand doesn’t push you; it bends the terrain so choosing it feels like going straight.

6. Trust is the binding energy. It converts loose attention into lasting structure — firms, institutions, currencies — and its loss vaporizes them.

7. Credibility is stored potential. A costly-to-build stock that acts at a distance and before you arrive, and compounds rather than depletes when used.

8. Common knowledge is the field. Coordination needs everyone to know that everyone knows; publicity is the medium through which social bodies feel each other.

9. To attend is to bestow mass. There is no neutral observation — watching a thing feeds it, so your attention is both weapon and gift.

10. Value is belief at critical mass. Money and valuations are collective belief gone self-sustaining; a bubble is a narrative self-gravitating past its limit.

11. Virality is gravitational collapse. Ignition rewards density and salience, not merit — which is why falsehood outruns truth.

12. Distribution beats product. Inside the band quality allows, the gravity well — not the artifact — sets the orbit.

13. You can copy the matter, never the gravity. Features are cheap and clonable; the well of accreted attention took years and cannot be reached by copying the object.

14. Social gravity is measurable. Influence has an exact geometry — PageRank, centrality, the h-index, the literal gravity models of trade.

15. Attention is the reserve currency. Money, status and power are banknotes; attention is the gold they are all convertible into, and it is upstream of all of them.

16. Consciousness has a center of mass. At scale, minds bind into one field — the noosphere — and the human universe increasingly exists inside consciousness, not the reverse.

1. Attention is the mass of the human world

Metaphor: Attention is the iron in the ore — the raw mass everything else is smelted from.

Definition:
Attention is the finite capacity of minds to notice, and it is the substance the social force couples to.
In a world flooded with information, attention — not data, money, or truth — is the one genuinely scarce input.
Every consequential act is downstream of an attention transaction that settled first.
Support, trust, belief, and money are simply attention that has condensed into denser forms.
It is conserved and rivalrous: a minute you give is one you can never give again.
Master this and the other fifteen principles are just descriptions of what attention does.

Why it holds:

• Simon: “a wealth of information creates a poverty of attention.”

• Total human attention is hard-capped at waking-minutes × population.

• Attention markets (Wu’s “attention brokers”) price it like a commodity.

• Every feed, ranking, and ad auction exists only to allocate it.

• Nothing scales without capturing it first — no exceptions found.

How to build with it:

• Budget attention like cash; it is your true P&L.

• Win a narrow slice completely before widening.

• Treat every feature as an attention cost, not only a benefit.

• Measure “share of the minds you care about,” not raw reach.

• Never spend attention you can’t convert into trust.

2. The weakest force builds the world

Metaphor: A fridge magnet beats the whole Earth — yet the Earth, not the magnet, holds the moon.

Definition:
Consciousness is trivially overridden in any single interaction and decisive across a population.
The “strong” forces of daily life — price, convenience, habit — win locally and cancel globally.
Attention and trust never cancel, so they alone accumulate to the largest scale.
No hard power survives the collective withdrawal of belief.
What feels like weakness is a scale error: you’re watching two grams, not the galaxy.
The force that looks negligible up close is the one that decides what stands.

Why it holds:

• Gravity is ~10^-39 the strong force yet governs galaxies.

• Charges neutralize; mass and attention do not.

• Regimes and currencies collapse when belief withdraws, regardless of arms.

• Aggregation, not intensity, is the source of the power.

• Long-range plus one-signed forces dominate at scale.

How to build with it:

• Don’t judge traction by any single “no.”

• Play for aggregate belief, not individual persuasion.

• Build for the population default, not the hero user.

• Expect slow-then-sudden: accumulation crosses a threshold.

• Guard the belief base; it’s the one thing you can’t buy back.

3. Attention only ever attracts — so it can only concentrate

Metaphor: There is no anti-gravity — and no anti-attention; even hatred pulls you closer.

Definition:
Attention has one sign: to notice something is to feed it, never to repel it.
Outrage, envy, and scandal are all inflows of mass, not outflows.
With no repulsive term, the system cannot spread out; it can only clump.
The stable outcome is a power law — extreme concentration — not an even spread.
“Bad” attention still raises your gravitational rank.
So the natural state of any attention field is radical inequality.

Why it holds:

• Power-law signatures recur across web, citations, wealth, fame (Newman).

• Controversy reliably increases reach, not decreases it.

• One-signed forces have no even-distribution equilibrium.

• Attempts to cancel often amplify the cancelled.

• Empirically, a tiny fraction of nodes holds most of the links.

How to build with it:

• Fear obscurity, not controversy.

• Convert any attention — even critical — into a next step.

• Prefer polarizing to being ignored.

• Design for concentration; pick the one node you want to be.

• Never optimize for “balanced” attention — it doesn’t exist.

4. Mass attracts mass — the rich get richer

Metaphor: Dust becomes a planet because the lump that starts bigger eats faster.

Definition:
New attention attaches preferentially to whoever already has attention.
Each increment of mass raises the pull, which gathers the next increment faster.
The result is runaway accretion: a few giants, endless debris.
Merton called it the Matthew effect — to those who have, more is given.
Early leads compound into structural, near-permanent dominance.
The distribution is not merit tilted by luck; it’s gravity by construction.

Why it holds:

• Barabási–Albert preferential attachment reproduces real networks.

• Citations, followers, and wealth all show cumulative advantage.

• Bestseller lists are self-reinforcing by design.

• Platform defaults get chosen because already chosen.

• Small early differences produce huge late gaps (path dependence).

How to build with it:

• Engineer a visible early over-density: launch concentrated, not diffuse.

• Make your current mass legible — show the numbers that attract more.

• Seed where accretion can start, not where it’s already saturated.

• Get the early lead; it compounds harder than late quality.

• Once ahead, reinvest attention into more attention.

5. Consciousness curves the space others move through

Metaphor: The star doesn’t tug the planet — it bends the road so “straight ahead” leads home.

Definition:
A massive brand doesn’t force a choice; it reshapes the landscape of choice.
It makes itself the low-friction default — the obvious, the “of course.”
People experience no coercion; they simply follow the tilted terrain.
Salience, not merit, is what accumulated attention manufactures.
Choosing the incumbent costs less cognitive energy, so minds roll toward it.
Rivals aren’t beaten on features — they’re asked to climb uphill.

Why it holds:

• Einstein: mass-energy curves spacetime; curvature routes motion.

• Bordalo–Gennaioli–Shleifer: we overweight the salient, not the best.

• PageRank literally ranks by the curvature of link-space.

• Defaults dominate choices in every studied domain.

• “What everyone uses” is a gradient, not a fact.

How to build with it:

• Become the default in one context, completely.

• Cut the cognitive cost of choosing you to near zero.

• Own the reference point (”the X of Y”) others get compared to.

• Plant yourself in the shared file, the standard, the tutorial.

• Make switching away feel like walking uphill.

6. Trust is the binding energy that turns attention into structure

Metaphor: Attention is gas; trust is the pressure that makes it a star that burns for a billion years.

Definition:
Attention attracts, but without binding it gathers and disperses.
Trust is the depth of the well that holds attention in a lasting form.
It’s what lets attention become a firm, an institution, a currency.
Trust is expensive to build, measurable in payoff, catastrophic to lose.
Remove it and the structure flies apart though its matter is untouched.
It is the whole difference between a crowd and an organization.

Why it holds:

• Coleman: social capital creates otherwise-impossible human capital.

• Knack–Keefer, La Porta, Algan–Cahuc: trust predicts and causes growth.

• Money and contracts run on trust, not enforcement alone.

• Edelman tracks institutions rising and falling with trust.

• Scandals vaporize century-old firms overnight.

How to build with it:

• Convert every burst of attention into a kept promise.

• Under-promise on the loud channel; over-deliver on the quiet one.

• Treat reliability as a growth strategy, not an ops detail.

• Protect trust like principal — slow to earn, instant to burn.

• Build rituals of consistency; repetition is what binds.

7. Credibility is stored potential — paid forward, never spent

Metaphor: Lift a weight and it holds energy; credibility is altitude that does work before you enter the room.

Definition:
Reputation is a stock you charged up with costly, hard-to-fake effort.
It acts at a distance and in advance of your actual presence.
Unlike money, using it doesn’t deplete it — it compounds.
It closes deals, raises rounds, and wins hires before you speak.
Only real, expensive signals can lift you to altitude.
The gap you feel to an incumbent is altitude, not features.

Why it holds:

• Spence: credentials work because they’re costly, not instructive.

• Zahavi/Grafen: the honest signals are the ones you can’t fake.

• Resnick–Zeckhauser: eBay reputation earns premiums on identical goods.

• Nowak–Sigmund: reputation sustains cooperation among strangers.

• Trusted names raise on a story where unknowns need a spreadsheet.

How to build with it:

• Spend real effort on visible, hard-to-fake proof.

• Bank credibility before you need it; it acts before you arrive.

• Reuse it freely — deploying it grows it.

• Stake something costly to make a claim believable.

• Never trade long-run credibility for a short-run spike.

8. Common knowledge is the field

Metaphor: Iron filings are random until one field snaps them all into the same line.

Definition:
People coordinate not on private belief but on mutually known belief.
What matters is everyone knowing that everyone knows, recursively.
Publicity — shared, simultaneous attention — is that field.
A fact known privately by all can still move no one.
Public moments manufacture the knowledge that enables joint action.
Revolutions and currencies turn on this recursion, not on private opinion.

Why it holds:

• Schelling focal points; Aumann’s formal common knowledge.

• Chwe: rituals and broadcasts exist to create common knowledge.

• Morris–Shin global games: public signals drive coordination.

• Bank runs and revolutions ignite on visibility, not new facts.

• A currency holds when each knows all others will accept it.

How to build with it:

• Engineer the public moment everyone sees everyone seeing.

• Prefer one simultaneous reveal to diffuse private reach.

• Make adoption visible — badges, counts, logos, crowds.

• Create rituals and landmarks that become shared reference points.

• Turn “many privately like us” into “everyone knows everyone does.”

9. To attend is to bestow mass — there is no neutral observation

Metaphor: Every glance is a vote; the crowd watching the fire is feeding it oxygen.

Definition:
Giving attention is giving support; noticing is never passive.
Even hostile or ironic attention adds to the observed thing’s mass.
There is no view from nowhere — the audience alters what it watches.
What you attend to, you strengthen, whether you endorse it or not.
Your attention is a resource others harvest by provoking it.
Withdrawal of attention, not opposition, is the real starve.

Why it holds:

• Engagement metrics reward reaction of any valence.

• Boycotts that trend frequently boost sales.

• Algorithms amplify what’s attended, not what’s approved.

• “Don’t feed the troll” encodes exactly this law.

• Attention, not agreement, is what the system counts.

How to build with it:

• Choose what you feed — your attention is both weapon and gift.

• Starve rivals of reaction instead of fighting them loudly.

• Design hooks that convert reflex attention into a step.

• Don’t amplify what you oppose by attacking it publicly.

• Ask of any outrage: do I want to give this mass?

10. Value is belief that reached critical mass

Metaphor: Money is a shared hallucination that enough people had at once for it to come true.

Definition:
A valuation is a fact about a crowd’s belief, not about an object.
Once enough consciousness pools, the belief becomes self-sustaining.
Money works purely because we jointly, bindingly believe it does.
Prices move on contagious stories more than on fundamentals.
A bubble is a narrative self-gravitating past its binding energy.
Value lives in the field of belief, not inside the thing.

Why it holds:

• Shiller: economies move on viral narratives.

• Keynes’s beauty contest; Soros’s reflexivity.

• Bikhchandani–Hirshleifer–Welch: rational cascades tip crowds.

• Fiat money, brands, and tokens have no intrinsic value.

• De Long et al.: sentiment, not value, moves prices.

How to build with it:

• Build and seed the narrative, not just the asset.

• Reach a critical mass of believers before optimizing the product.

• Watch for reflexive loops you can ride — or that can turn.

• Make belief legible so it recruits more belief.

• Respect that belief bound too fast can collapse.

11. Virality is gravitational collapse — and it doesn’t check quality

Metaphor: A gas cloud doesn’t ask to be good before it ignites — only to be dense enough.

Definition:
Ideas condense from noise by a threshold instability, like star formation.
Ignition rewards density, novelty, and emotional charge — not merit.
Many ideas need several reinforcing contacts to catch (complex contagion).
Below the threshold, nothing; above it, runaway spread.
Falsehood often outruns truth because it recruits attention better.
To spread an idea you engineer over-density, not more correctness.

Why it holds:

• Vosoughi–Roy–Aral: false news spreads farther, faster, deeper.

• Centola: complex contagion needs multiple exposures.

• Goel et al.: structural virality is a true chain reaction.

• Bettencourt: ideas spread on epidemic mathematics.

• Emotional, novel content wins attention reliably.

How to build with it:

• Optimize for shareable density, not comprehensiveness.

• Seed clusters where reinforcement is likely, not lone nodes.

• Engineer novelty and emotional charge deliberately.

• Cross the threshold in one place before broadening.

• Make retransmission effortless and status-giving.