Prepared by the ENSI Foresight Division. Built on the 183-document primary library assembled for this programme, with no new sourcing. Compiled 24 October 2026.
The argument, before the list
There is a paradox at the centre of the specialised-unit economy, and it is not a quibble at the edges. It is structural, it compounds, and nobody has a solution.
The arrangement makes tacit judgement the scarcest input in the economy while dismantling the only known production route for it.
Here is the mechanism. The functions that package most cleanly into performance-valued units are the ones with attributable output, a clean handover boundary and bounded error cost — lead generation, sourcing and screening, bookkeeping and close, receivables, first-line support, test authoring, first-pass contract review. That list is also, almost exactly, the list of junior roles. It is how people have always learned a trade: by doing the detailed work for several years, under supervision, accumulating the local, tacit, fast-changing knowledge that later makes senior judgement possible.
And that knowledge is precisely what the whole arrangement runs on. The strongest theoretical argument for a person operating an agent stack is not that the person works faster. It is that Hayek’s objection was never answered — the computational problem was solved decades ago, with sparsity reducing a ten-million-product allocation from roughly 10²¹ operations to about seventeen minutes, while the acquisition of dispersed local knowledge remains exactly as hard as it was. The operator’s advantage is that they hold coefficients the system does not have. Remove the years of detailed work and you remove the acquisition process.
This is not a prediction that firms will fire their juniors. The evidence says the opposite about the mechanism: the OECD found no measured employment or wage effect from AI adoption, with adjustment through slowed hiring rather than dismissals. Which is worse for formation, not better. A firing wave would at least release experienced people who already hold the coefficients. A hiring freeze quietly closes the entrance while everyone inside keeps their position, and the damage shows up in a decade as an absence rather than a crisis — no cohort of people who understand how the packaged work actually behaves.
So the question this report answers is narrow and urgent: if the detailed work is gone, where do the coefficients come from, and what should talent development become instead?
Three findings shape the answer, and two of them are uncomfortable.
First, nobody will pay for formation under the default arrangement, and this is a structural result rather than a failure of will. In the firm era the employer paid for training because tenure let it capture the return. A supplier economy removes tenure, and in a one-person unit “its people” is one person who is the unit. The classic human-capital problem — that the investor cannot capture the return because the trained person can leave — becomes acute. There is a known institutional answer to exactly this, and it comes from the coordinated economies: collaborative vocational training, in which employers fund formation collectively because no single employer can capture the return. That template, not exhortation, is the move.
Second, assessment in this arrangement will be mostly noise, and will be treated as signal. This is the most important transferable finding in the library and it comes from an unexpected place. Working across roughly 50,000 subjects, Marek and colleagues established that brain-wide association effects have a median of |r| = 0.01, a top-1% threshold of |r| = 0.06, and a maximum replicated value of |r| = 0.16 — so at the typical published sample of about 25 people, the 99% confidence interval is r ± 0.52, wide enough for two honest laboratories to reach opposite conclusions from sampling variability alone. Their finding was that reproducible inference required thousands of subjects.
Now transpose the statistics, because the structure of the problem is identical. A performance-valued unit two years old has perhaps twenty completed engagements. That is n = 20. The effect sizes that distinguish a good specialist from an average one are not large, the outcomes are confounded by client quality, market conditions and luck, and the sample is tiny. An axiocracy that reads early unit performance as a measure of ability is making exactly the inference Marek showed cannot be made — and because the companion report argues reputation should be portable, the noise becomes permanent. Initial conditions get converted into durable standing. That is not a meritocracy; it is a lottery with a memory.
Third, the skill regime you choose determines the kind of innovation you get, and the supplier economy picks a side by accident. The comparative-institutions evidence is specific here. Economies built on general, portable skills and fluid labour markets specialise in radical innovation; economies built on firm-specific skills, long tenure and collaborative training specialise in incremental innovation. The measurement is clean: across 30 technology classes in European Patent Office data, Germany’s specialisation profile is “almost the mirror image” of the United States’, stable across a decade — Germany concentrated in mechanical engineering, product handling and transport, the US in classes characterised by radical innovation.
A supplier economy of portable specialists is, by construction, the general-skills configuration. So expect more radical innovation and less incremental capability — more new things, less compounding refinement of existing ones. That is a real trade, it is rarely stated, and it is probably not what anyone intends to choose.
One encouraging finding before the lists, because the picture so far is bleak and the evidence does not actually support despair. Plasticity is real and deficits are not uniformly permanent: in the cleanest natural experiment on severe early deprivation available, children institutionalised for over six months showed persistent elevated symptoms into their mid-twenties — and yet cognitive impairment fully remitted by adulthood (p = 0.76), with a fifth of the severely deprived group problem-free at every assessment. Formation is recoverable far later than the pessimistic reading assumes. That is not a licence to destroy the entry ladder. It is a reason to think a rebuilt one can work on people who missed the first.
The main points, in brief
The work being packaged first is the work that formed people. Tier-one functions — attributable, interface-clean, bounded error — are the junior roles, and removing them removes the route by which anyone becomes senior.
A hiring freeze is worse for formation than a firing wave, because it closes the entrance without releasing anyone who already holds the knowledge.
Nobody will fund formation under the default arrangement. Tenure was what let the employer capture the return; a supplier economy removes it. The known answer is collective sectoral funding, not goodwill.
Early assessment will be noise treated as signal. A two-year-old unit has roughly twenty engagements. At that sample size, on effects this small, the confidence interval swallows the inference — and portable reputation makes the error permanent.
The skill regime decides the innovation type. General portable skills produce radical innovation; firm-specific skills produce incremental. Germany’s patent profile is “almost the mirror image” of America’s across 30 technology classes. A supplier economy picks radical by accident.
Formation shifts from knowing to framing. When execution is absorbed, the scarce skill becomes selecting which problem to solve and specifying it well enough to be verified.
Attribution literacy becomes a core competence, not a finance skill — knowing whether your own number means anything is a survival requirement in a system that pays against numbers.
Self-authored formation outperforms imposed curriculum, on the same evidence that self-chosen sanctioning reached 90% of optimal returns while externally imposed rules with weak enforcement increased defection.
Depth now requires explicit edges. A specialist who cannot say where their competence ends is dangerous in a configuration where nobody else is checking.
Public transition capability is an order of magnitude short. Denmark spends 1.2% of GDP on active labour market policy against an OECD average of 0.4% and a UK figure of 0.3% — and even Denmark’s protections reach none of the solo self-employed.
How this report is organised
Part One sets out the six shifts in where formation effort goes — what talent development stops optimising for and what it starts optimising for.
Part Two names the eight attributes that become critical, each with what it replaces and how it can actually be built, plus the four things that depreciate and should stop being trained.
Part Three sets out what has to change institutionally for any of this to be produced at scale, including the funding question that determines whether the rest is possible.
A note on evidence. No document in this library studies the formation of agent-operated specialists, because they barely exist. What the library contains is strong evidence on the four things the answer depends on: how skill regimes shape economies, what sample sizes are required to make inferences about individuals, what happens when training investment cannot be captured by the investor, and whether imposed or self-authored rules produce better outcomes. Where this report reasons by transfer — chiefly from the neuroimaging statistics to performance assessment, and from collaborative vocational training to supplier formation — the transfer is named so it can be challenged.
Part One — The six shifts in where formation effort goes
1. From breadth-then-depth to depth plus rented breadth
What changes. The standard career architecture has been breadth first, depth later: learn the business, rotate, acquire general competence, then specialise once you have enough context to choose well. That sequence existed because of a constraint that has now lifted. A narrow specialist could not operate alone — someone had to supply the administration, the client acquisition, the adjacent competencies — and the firm supplied them in exchange for setting the minimum scope at which a person could function.
Agents supply the adjacent competencies directly, so the minimum viable scope collapses and depth becomes available far earlier than the old sequence allowed. Breadth does not stop mattering; it stops needing to be resident. It is rented, from the stack and from other units.
What this means for formation. Stop treating the rotation as the foundation and start treating it as context acquisition with a deadline. The thing that genuinely requires years is the accumulation of local tacit coefficients inside a specific problem domain, and that accumulation starts later under the old model than it needs to.
The honest risk, and it is the mirror image of the opportunity: depth without accumulated context produces specialists who are precise and wrong — fluent inside their domain and unable to see that the problem they were handed was the wrong one. Which is why shift two matters more than this one.
2. From knowing to framing
What changes. When execution is absorbed, the binding constraint moves upstream to problem selection and specification. The skill is no longer producing the answer but deciding which question is worth answering and stating it precisely enough that the answer can be verified.
This is the shift that reallocates scarce judgement, and it is the one most visible to anyone currently working this way. People who are good at framing have historically spent most of their hours on work that did not require framing, because somebody had to do it and they were there. Removing that work does not make them better at framing; it gives them more hours in which to do it, which is a different and larger gain than a productivity improvement.
What this means for formation. Problem framing has been treated as something that emerges from experience rather than something taught, and that was affordable when there were twenty years of detailed work in which it could emerge. It is no longer affordable. The components are teachable and rarely taught: distinguishing a symptom from a cause, stating a problem in terms that admit a falsifiable answer, identifying what would have to be true, and — most neglected — knowing when to decline a problem as badly posed.
The evidence that this is the durable half of the work comes from the same place as everything else in this programme: the information asymmetry that favours the operator is about acquisition of local knowledge, not processing. Framing is acquisition work. It is the part that does not transfer to the system as models improve.
3. From execution quality to attribution literacy
What changes. In a system that pays against measured contribution, the ability to tell whether your own number means anything stops being a finance competence and becomes a survival requirement.
The reason is that attribution and measurement are routinely confused, including by sophisticated people. The cautionary case in this library is instructive precisely because it comes from professional economists: the headline corporate-markup series that framed a decade of policy debate implied roughly 70% of GDP as pure economic profit and a capital output elasticity of −0.25, and an independent reconstruction using a fuller cost measure found no rise at all. Confident numbers, wrong object. If that can happen in peer-reviewed macroeconomics, it will happen constantly in a quarterly unit scorecard.
What this means for formation. Three specific competences, none of which currently appears in professional training outside analytics roles.
Knowing what your sample can carry. This is the Marek lesson and it generalises exactly. At small samples, with small true effects, confidence intervals swallow the inference — the median brain-wide association effect was |r| = 0.01 and at n = 25 the 99% interval was r ± 0.52. A specialist reporting a quarterly improvement on twelve engagements needs to know that they have not measured anything, and needs to be able to say so without it reading as an excuse.
Distinguishing attributable from merely measurable. Plenty of functions emit abundant numbers that cannot be causally assigned to the unit producing them. Being able to name the confounder in your own favour is a professional obligation in this arrangement.
Contesting a scoring function. Not gaming it — contesting it, on the record, with an argument. The companion report makes the case that scoring functions are distributive institutions pretending to be measurement ones: identical data produces either near-equal payouts or a Gini around 80 with a 2.5-to-1 median gender gap depending on a weighting nobody voted on. Someone has to be able to see that and say it.
4. From managing people to directing agents and carrying consequence
What changes. The classical promotion path converts a good practitioner into a manager of practitioners. In a unit of one operator plus a stack, the equivalent skill is direction and verification — specifying work precisely, checking output you did not produce, designing the escalation path for the cases the stack handles badly, and holding the consequence when it is wrong.
The last part is the one that gets skipped, and it is the whole difference between a unit and a vendor. Accountability is the thing the firm is actually buying. A unit that produces output without carrying the consequence of its failure has not been given ownership; it has been given a task.
What this means for formation. Two things almost nobody is trained in.
Verification of work you did not do. This is a distinct skill from doing the work, and it is closer to editing or auditing than to production. It requires knowing the characteristic failure modes of the producer — which, for an agent stack, are different from a junior colleague’s and change with each model revision.
Designing the appeal path. The gig evidence shows exactly what happens when this is absent: chat support that is the most-used feature and fails precisely on penalties, deactivations and payouts; workers penalised by verification systems that misfire; drivers deactivated after complaints about accidents they did not cause. An operator who builds an agent-run process without a time-bound route to a human has built the thing the evidence says does the damage — and will eventually be on the receiving end of someone else’s version of it.
5. From firm-specific capital to portable capital — and the trade nobody is choosing deliberately
What changes. Human capital that only has value inside one organisation is worthless to a unit that needs to serve several. So formation shifts toward portable, general, transferable competence — and that is not a free improvement. It is a choice with a measured cost.
The comparative-institutions evidence is unusually specific. Economies built on general portable skills and fluid labour markets specialise in radical innovation. Economies built on firm-specific skills, long tenure and collaborative vocational training specialise in incremental innovation. The measurement: across 30 technology classes of European Patent Office data, Germany’s specialisation profile is “almost the mirror image” of the United States’, stable across 1983–84 and 1993–94 — Germany concentrated in the classes characterised by incremental innovation, the US in those characterised by radical.
And the mechanism is complementarity, which means it cannot be mixed and matched: “two institutions can be said to be complementary if the presence (or efficiency) of one increases the returns from (or efficiency of) the other.” General education complements fluid labour markets, which complement finance that depends on current profitability, which complements top-management-dominated decision-making. The parts come as a set.
What this means for formation. A supplier economy of portable specialists is the general-skills configuration by construction, so expect more new things and less compounding refinement of existing ones. Anyone whose competitive position rests on incremental capability — precision manufacturing, process industries, anything where the advantage is twenty years of accumulated small improvements held in people who stayed — should treat the supplier model as a threat to that advantage rather than a neutral efficiency gain. The mitigation is not rhetorical: it is retaining firm-specific formation deliberately in the places where incremental capability is the business, and accepting the cost of doing so.
6. From curriculum to self-authored formation with accountable monitoring
What changes. The default institutional response to a skills problem is a curriculum — a defined programme, imposed from above, with completion as the measure. The evidence says this is the weaker design, and the evidence is strong.
Subjects who chose their own sanctioning system after discussion achieved 90% of optimal net returns — the best result in any common-pool-resource experiment. An externally imposed limit with a one-in-sixteen inspection probability increased defection relative to discussion alone. And across 91 studies of self-governing arrangements, monitor accountability was the strongest of eleven design principles, with a supportive-to-unsupportive ratio of 11.7. The general finding: imposed rules with weak enforcement produce worse outcomes than no imposed rules at all, because they displace something that was working.
There is a cautionary case on the formation side specifically, and it comes from the exemplar. At one Mondragón plant, workers bought membership at €13,000 each, almost 80% joined for job security, and only 7% could state the group’s cooperative principles a year later. The formation failed inside the organisation most committed to it — which suggests that transmitting an operating philosophy is harder than transmitting a skill, and that assuming it happens by immersion is a mistake.
What this means for formation. Let units and individuals author their own development objectives and their own evidence of progress, with the institution holding a veto rather than the pen, and with whoever assesses them accountable to them in some concrete way. Completion of an imposed programme is the metric most likely to be satisfied without the capability being acquired.
Part Two — The eight attributes that become critical
Each is given with what it replaces, why it matters in this specific arrangement, and how it can actually be built — because an attribute with no production route is a wish list entry.
1. Problem framing
Replaces: domain knowledge as the primary asset.
When execution is absorbed, the value moves to deciding what is worth doing and stating it so the answer can be checked. This is the single highest-return attribute in the set and the one least deliberately produced.
How to build it. Volume of framed problems with feedback on the framing rather than on the answer — which means someone senior has to review the question and say why it was the wrong one. That review is cheap, fast, and almost nobody does it. The thing being trained is the ability to decline a badly posed problem, which is also the thing most likely to be punished in a junior role.
2. Judgement under thin evidence — and knowing when the evidence is thin
Replaces: confidence as a proxy for competence.
This is the attribute the measurement literature argues for most directly, and it has two halves that are usually conflated. The first is acting well when the evidence is insufficient, which is most of the time. The second, rarer and more valuable, is knowing that it is insufficient — and the quantitative benchmark for how bad human intuition is here is the finding that at n = 25, with true effects around |r| = 0.01 to 0.06, the 99% confidence interval is r ± 0.52. Two honest observers reach opposite conclusions from noise. People do not intuit this; they have to be taught it.
How to build it. Make people state a prediction with an interval before they see an outcome, and then score the interval rather than the point. This is the only reliable method for producing calibration, it is well established, and it is almost never used outside forecasting communities.
3. Attribution literacy
Replaces: reporting skill.
In a system that pays against measured contribution, the ability to say whether a number is causally yours becomes a core professional competence rather than an analyst’s specialism. The discipline is specific: distinguish attributable from merely measurable, name the confounder that flatters you, and resist the composite score — on the principle that combining distinct quantities into one number is like a car meter “that added up in one single number the current speed of the vehicle and the remaining level of gasoline.”
How to build it. Have people argue against their own best number, in writing, as a routine part of reporting. The practice is adversarial and the output should be kept, because a track record of having caught your own false positives is itself a strong reputational asset in an arrangement full of people who have not.
4. Agent direction and verification
Replaces: people management as the promotion path.
Specifying work precisely, checking output you did not produce, knowing the characteristic failure modes of the producer, and designing the escalation path for the cases handled badly. Verification is nearer to editing or auditing than to production, and it is a distinct skill from doing the work.
How to build it. Deliberate exposure to known-bad output — review exercises where the errors are planted and the trainee’s score is the catch rate. This is standard in auditing and radiology and essentially absent from the rest of professional training. It also has to be refreshed, because the failure modes of a stack change with each revision in a way a junior colleague’s do not.
5. Accountability tolerance
Replaces: escalation as the default response to risk.
Carrying the consequence is what distinguishes a unit from a vendor, and it is the attribute most likely to be the real constraint on how many people can actually do this. Owning a process means the bad outcome is yours in public. A great many highly capable people have organised their careers specifically to avoid that, which is a rational response to how organisations have historically treated visible failure.
How to build it. Graduated exposure with a survivable downside — real ownership of real outcomes at a scale where being wrong is instructive rather than career-ending. This is also where the floor from the companion report does work that looks economic and is actually developmental: a credible floor is what makes the downside survivable enough to accept. The Kenyan evidence is that people invested in enterprise because income was guaranteed for twelve years, and that the gain was lost to a single shock — read as a formation finding rather than a welfare one, that says risk-tolerance is manufactured by durable protection, not by exhortation.
6. Narrow depth with explicit edges
Replaces: generalist adaptability.
Depth is now available earlier and is more valuable, but it comes with an obligation that generalism did not impose: knowing precisely where your competence ends. In a firm, the edges of your competence were patrolled by colleagues and process. In a unit, nobody is checking. A specialist who cannot name their own boundary is dangerous in exact proportion to how good they are inside it.
How to build it. Require an explicit statement of scope — what the unit does, what it does not do, and what it refers onward — as a condition of engagement, and treat a referral as a success metric rather than a lost sale. The commons evidence supports this structurally: the strongest design principle across 91 studies was monitor accountability, and the second-order finding was that arrangements work when boundaries are clearly defined. Undefined boundaries are the failure mode, not a sign of flexibility.
7. Trust production
Replaces: the employer’s name as the credential.
A specialist inside a firm borrows the firm’s trustworthiness. A unit has to manufacture its own, and the infrastructure for doing so does not exist — the companion report flags a portable, verifiable, gaming-resistant, contestable delivery record as a genuine open problem whose four properties are in tension.
Until it exists, trust is produced the slow way: by a legible track record, by referral, and by being the person who said the uncomfortable thing early. The structural point worth knowing is that this is a constructed institution rather than a natural fact — credentials, licensure, limited liability and the enforceability of a reference are all legal artefacts that currently presuppose employment, which is precisely why they can be rebuilt.
How to build it. Document delivery contemporaneously and in a form a counterparty can verify, from the first engagement. The cost of starting this late is that the early record — the period when the sample is smallest and the reputation most malleable — is the part that cannot be reconstructed.
8. Taste
Replaces: correctness as the terminal standard.
When a stack can produce many acceptable outputs cheaply, the scarce act is selection among them. Taste is the compressed, mostly tacit judgement about which acceptable answer is the right one for this client, this moment, this constraint — and it is the attribute least amenable to instruction, which is why it will be the most durable source of advantage.
How to build it. Exposure to large volumes of worked examples with the reasoning attached, and critique from someone whose judgement is already formed. This is how it has always been produced in design, law and medicine, it does not appear to have a shortcut, and it is the strongest remaining argument for apprenticeship as a form — which is awkward, because apprenticeship is what the entry-ladder problem is removing.
The four things that depreciate, and should stop being trained
Breadth held as insurance. Knowing a little about everything was a hedge against the firm reassigning you. In a unit, rented breadth is cheaper and better than resident breadth, and hours spent maintaining shallow coverage are hours not spent accumulating depth. Breadth as context for framing remains essential; breadth as fallback capability is now a cost.
Process compliance as a competence. Following a defined procedure accurately is the thing that automates most completely and first. Training people to execute a process correctly is training them for the tier-one work that is leaving. What survives is the ability to design the process and to notice when it is producing the wrong result.
Coordination as value added. A large share of senior roles consist of moving information between people who could not easily reach each other. That function is being compressed from both ends — by the stack and by smaller units with fewer handoffs. Coordination remains necessary and stops being a career.
Credential signalling. In an arrangement that pays against attributable output, a credential is a prior that gets overwritten by the first ten engagements. It does not become worthless — it still opens the first door, and for regulated activity it remains mandatory — but it stops being the asset and becomes the entry ticket. The caution from the capability literature is worth carrying here: conversion factors mean identical resources yield unequal real freedom, so identical training will produce unequal capability, and a system that treats the credential as the measure will mis-rank people systematically and with confidence.
Part Three — What has to change institutionally
The attributes in Part Two do not get produced by individuals deciding to acquire them. Formation has always been an institution, and the institution that supplied it is being dismantled. Six changes, ordered by whether the mechanism is known.
1. Collective sectoral funding — the one change everything else depends on
The problem, stated structurally. In the firm era the employer paid for training because tenure let it capture the return. A supplier economy removes tenure. In a one-person unit, “its people” is one person who is the unit, and that person cannot borrow against their own future competence. The investor and the beneficiary have been separated, which is the classic human-capital problem in its most acute available form.
Exhortation will not close it, and neither will individual responsibility, because the shortfall is not a failure of will. It is a failure of appropriability.
The known answer. Coordinated economies solved precisely this with collaborative vocational training — employers funding formation collectively because no single employer could capture the return, supported by dense business associations and industry-level coordination. That is the mechanism, and the library also supplies a working financing structure to copy: Italy’s worker-buyout fund is built on 3% of the cooperative sector’s annual surplus, matched one-to-one against workers capitalising up to three years of advance unemployment benefit, producing 257 buyouts, roughly 9,300 jobs, and survival at 13.12 years against a national firm average of 13.5 from a distressed starting point.
Apply that structure to formation rather than acquisition: a sectoral formation fund financed by a levy on the firms buying from units, drawn on by units to train. The firms buying the output are the ones capturing the return, which makes them the right payers — and a sectoral levy solves the free-rider problem that stops any individual buyer from doing it alone.
Status: mechanism known, not yet applied to this. This is the most important item on the list and the most tractable, because it is a translation rather than an invention.
2. Procurement conditioned on formation
The move. A firm buying from external units requires, as a contract term, that the units train — in the same way states have long conditioned public contracts on apprenticeship provision. The precedent is small and real: procurement conditioned on apprenticeship training appears in the predistribution toolkit as an established instrument.
Why it works where exhortation does not. It attaches the obligation to the thing the firm wants, which is the output. And it reaches the population that most needs it: a unit under competitive pressure will not train voluntarily, because training is a cost whose benefit the next client may capture.
Status: available now, needs no legislation. Any large buyer could put this in its standard terms next quarter. It is the single most implementable item in this report.
3. A deliberately retained training surface
The problem. If tier-one work leaves entirely, the acquisition process for tacit coefficients leaves with it. The formation route and the automation target are the same work.
The move, and it is deliberately uncomfortable. Retain some of the packaged work inside the firm, at some cost to efficiency, specifically as a formation surface — and account for it as such rather than pretending it is operationally necessary. A firm that automates 100% of its tier-one functions has optimised a cost line and removed its only internal supply of people who understand how that work behaves. The right number is not zero and is probably not large.
The supporting evidence for taking the cost seriously: the comparative-institutions finding that firm-specific skills, long tenure and collaborative training complement incremental innovation means that organisations whose advantage rests on accumulated refinement are the ones with the most to lose from full externalisation. For them, retaining a training surface is not sentiment; it is protecting the mechanism that produces their actual competitive position.
Status: known, resisted. The obstacle is that the cost is visible and immediate while the benefit is invisible and arrives in a decade — which is the standard shape of an under-provided good.
4. Assessment standards that respect sample size
The problem. This is the most consequential and least obvious item, and it follows directly from transferring a statistical result. Early-unit performance records are tiny — twenty engagements is n = 20 — the true differences between specialists are modest, and the outcomes are confounded by client quality and luck. The reference finding establishes that at such samples, with such effects, the inference cannot be made: a median effect of |r| = 0.01, a maximum replicated value of |r| = 0.16, and at n = 25 a 99% interval of r ± 0.52. Reproducibility required thousands of observations.
An axiocracy will nonetheless read early performance as ability, and because the companion report argues reputation should be portable, the noise becomes durable. Initial conditions get laundered into permanent standing. That is the mechanism by which a system that calls itself a meritocracy produces a hereditary one in a single generation of units.
The move. Three concrete standards. Publish confidence intervals, not point scores, on any unit rating. Require a minimum engagement count before a rating is treated as informative, and display the count prominently where the rating appears. And make early records expirable or re-settable, so that a bad first year under a bad client is not a permanent sentence — the plasticity evidence supports this directly, since the cleanest natural experiment on severe early deprivation found cognitive impairment fully remitted by adulthood (p = 0.76) with a fifth of the severely affected group problem-free throughout.
Status: technically trivial, commercially unattractive. Every platform incentive points toward a single confident score.
5. Public transition capability at something like the right scale
The problem. Formation at national scale is a public capability, and the current provision is an order of magnitude short of the demonstrated requirement.
The benchmark. Denmark spends 1.2% of GDP on active labour market policy, against an OECD average of 0.4% and a UK figure of 0.3%. And the mechanism that makes the Danish model work is not benefit generosity: the administrative evidence shows displaced workers losing only 12–15% in the first year and 5–8% thereafter, with those small losses coming from fast re-employment and a compressed wage structure rather than from the size of the payment.
The caution, which is severe. Even Denmark — 67% union density, 82% bargaining coverage against the UK’s 27%, on-site representatives in one firm in two — leaves the solo self-employed and zero-hours workers outside the model entirely, with one in three workers now in non-standard employment. So the benchmark for spending is available and the benchmark for coverage is not. Nobody has built transition capability for a population of micro-suppliers.
The hopeful finding. Institutional quality appears to substitute for social trust: above roughly German levels of regulatory quality, trust’s marginal effect becomes statistically indistinguishable from zero. A low-trust state is therefore not condemned to fail at this — it has to build the institution rather than wait for the culture, which is a thing a government can actually do.
Status: funding known, coverage open.
6. The operator apprenticeship
The problem. The one attribute with no shortcut is taste — selection among acceptable outputs — and the only known production method is exposure to volume with the reasoning attached, under critique from someone whose judgement is already formed. That is apprenticeship, and apprenticeship is what the entry-ladder problem removes.
The move, stated honestly as untested. Attach entrants to operators rather than to firms — the unit as the formation site, with the apprentice doing framing and verification work alongside someone who already holds the coefficients. This has the right shape: it puts the learner next to the judgement rather than next to the detail, which is where the scarce knowledge now sits.
It also has two problems nobody has solved. The unit has no incentive to train, since the apprentice can leave and the next client captures the benefit — which is why item 1 is the precondition for this one rather than a complement to it. And it is unclear whether framing and verification can be learned without having done the detailed work, or whether the coefficients are only acquirable the slow way. That is the open empirical question at the centre of this entire report.
Status: open. The plausible answers — simulation against historical cases, deliberately retained manual work, apprenticeship to an operator — are untested, and the honest position is that we do not know which works.
What this adds up to
The arrangement forecast in this series makes tacit judgement the scarcest input in the economy and removes the only known route for producing it. That is the finding, and nothing in the evidence base dissolves it.
What the evidence does supply is a clear separation between the parts of the response that are known and the parts that are not.
Known, and mostly a matter of translation. The funding mechanism for formation when no single buyer can capture the return — collective sectoral provision, on the model of collaborative vocational training, financed on the structure of a sectoral levy that already works for worker buyouts. Procurement conditioned on training, which any large buyer could adopt immediately. Assessment standards that respect sample size, which are technically trivial. A public transition capability at roughly three to four times current typical spending, with a demonstrated benchmark.
Open, and the frontier. Whether framing and verification can be acquired without first doing the detailed work. Whether a formation record can be made portable, verifiable, gaming-resistant and contestable at once. And how to extend any protective or developmental institution to a population of micro-suppliers, which not even the best-covered labour market in the world has managed.
The most useful conclusion is about sequence rather than content. Of the six institutional changes, one is the precondition for the rest: unless formation is collectively funded, every other item fails on appropriability, because the party who would pay cannot capture the return. A firm will not train a unit that can leave. A unit will not train a person who can leave. An individual cannot borrow against competence they do not yet have. That is not a cultural problem and it will not be solved by anyone deciding to take formation seriously. It is a financing structure, the template exists, and it has been working in Italian cooperative law for forty years.
And one closing observation that cuts against how this subject is usually discussed. The attributes that become critical — framing, calibrated judgement, attribution literacy, verification, carrying consequence, knowing your edges, producing trust, taste — are almost exactly the attributes that have always distinguished senior professionals from junior ones. The shift does not require new human capacities. It requires producing the existing senior ones much earlier, in far more people, without the twenty years of detailed work that used to produce them. That is a harder problem than inventing a new skill, and it is a more tractable one than it sounds, because at least we know what the target looks like.