The Tragedy of the Cognitive Commons

Sources#

• AI and Job Postings: From Destruction to Creation?
• AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It
• Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think
• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence
• Deploying AI from pilot to production: A practical blueprint for CIOs and technical leaders
• Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants
• DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501
• How Does AI Change Labor Demand? Evidence from 41 Countries
• The Early Impacts of AI on Employment among Recent College Graduates
• The Human-AI Substitution Principle: When will you be replaced by AI in your organization?
• The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise
• When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration

Summary#

A conceptual framework arguing that the pool of validation-capable expertise in a profession is a commons, that its regeneration mechanism is entry-level work, and that AI adoption removes that mechanism through decisions no single organization has any reason to avoid. Lovett's contribution is not new data — it is a structure that makes an already-observed pattern legible as a collective-action failure rather than a labor-market adjustment.

The load-bearing pair of constructs:

• Internalized Mastery — deep domain knowledge, schemas, diagnostic judgment, pattern recognition. Forms only through sustained cognitive struggle on progressively harder problems. "The cognitive struggle is not incidental friction to be engineered away; it is the mechanism."
• Distributed Mastery — fluency in orchestrating AI to produce professional-quality output: prompting, curation, human-AI workflow design. Achievable without deep domain internalization.

And the mechanism connecting them, the Validation Tether: effective Distributed Mastery depends on Internalized Mastery, because catching a wrong-but-plausible AI output requires knowing the domain independently. Split into two levels that the framework insists are qualitatively different:

	What it checks	What it needs
Surface validation	Internal coherence, obvious errors, basic plausibility	Learnable from working with AI itself
Substantive validation	Domain-specific wrongness, contextually inappropriate but technically correct output, subtle flaws under a plausible surface	Internalized Mastery — not otherwise obtainable

Evidence note. practitioner-opinion — and the tier matters more than usual here. This is a conceptual paper: peer-reviewed (Sage's Human Resource Development Review, accepted 2026-07-06) but containing no original measurement. Every number below is secondhand, and is attributed to its original study rather than to this paper. Lovett is unusually disciplined about this himself — see the evidential ladder below.

The paper's own evidential ladder#

Rare enough to be worth recording as a method: the paper sorts its own claims into three tiers and asks readers to keep them apart.

1. Empirically grounded (published evidence, cited): cohort-specific early-career employment decline in AI-exposed occupations; AI-assisted performance gains that are access-dependent rather than transferable; widespread reduction in active validation of AI output.
2. Theoretically derived, indirectly corroborated: that these produce commons-like collective-action failure; that Distributed Mastery depends on Internalized Mastery.
3. Extrapolatory, untested: that depletion will show up as measurably reduced substantive validation across cohorts; that it will vary by the five-factor model; that governance can moderate it.

The paper states plainly that "the profession-level depletion this paper describes is a structural prediction … not an observed outcome." Read this page the same way: the dissociation is measured, the tragedy is a forecast.

Why expertise behaves like a commons#

Three structural properties, and the third is the interesting one:

1. Collectively dependent — every organization hires from a shared pool it did not fully train. A bank's AI risk model needs human validators; hospital safety needs clinicians who kept their diagnostic skills; AI-generated code needs engineers who can audit it.
2. Non-exclusive — a firm that eliminates junior roles still benefits from profession-wide expert capacity: it hires seniors trained elsewhere, consults specialists, and relies on regulators who are themselves domain experts. Classic free-riding (Olson 1965).
3. Degradable through regeneration failure, not visible depletion — and unlike a fishery, there is no bare hillside. "Existing experts continue validating AI outputs and managing exceptions competently, creating an impression of stability while the regeneration mechanism experiences disruption."

The overgrazing arithmetic: each organization that cuts an entry-level role "captures 100% of the efficiency gains … while distributing the expertise depletion cost across all organizations." The firm that keeps its pipeline is the one at a competitive disadvantage.

The accidental-alignment argument#

The sharpest idea in the paper, and the one most portable outside HRD:

"The prior equilibrium was not the product of governance; it was the product of accidental alignment. Organizations maintained developmental pipelines not because they recognized any stewardship obligation … but because they needed entry-level workers to perform entry-level work. The operational necessity of junior labor was the hidden governance mechanism."

Standard human-capital theory (Becker 1964) already predicted chronic under-investment in general training, since trained workers defect to competitors. What kept the system running anyway was that the training happened as a by-product of work that had to be done by someone. AI removes the work, not the training budget — so this is a second-order market failure: not under-investment in development, but elimination of the productive activity through which development happened for free.

The corollary is the uncomfortable one: nobody has to decide to stop maintaining the commons for it to stop being maintained.

Two mechanisms, and why only one is measurable#

	Mechanism 1: direct position elimination	Mechanism 2: augmentation without internalization
What happens	AI performs entry-level tasks; the role is cut	The role survives; the junior hits senior-level output via AI while skipping the struggle
Visible in	Payroll and headcount data	Nothing — "employment numbers may appear healthy while regeneration quality silently degrades"

This asymmetry is the framework's most useful practical claim. The labor-market evidence everyone cites captures only Mechanism 1. Mechanism 2 is invisible to every metric currently collected, which is precisely why the paper argues for no-AI performance assessment and error-detection audits (see the research agenda below).

The borrowed evidence, attributed#

Every figure here belongs to the cited study, not to Lovett:

• Brynjolfsson, Chandar & Chen (2025), Stanford Digital Economy Lab — payroll data covering 25M+ US workers: workers aged 22–25 in the most AI-exposed occupations fell 16% in relative employment (Oct 2022 – Sept 2025) while 35–49-year-olds in the same occupations grew 8%+, with less-exposed occupations flat across all ages. Declines concentrate where AI automates rather than augments. (Figure superseded 2026-09-22: the vault now holds the paper itself — Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, revised August 2026 with data through June 2026 — and the number Lovett quotes secondhand is two vintages old. The kept-pace shortfall is 19%; see the section below for the full reconciliation. Lovett's use of the finding is unaffected: it moved in the direction his argument predicts.)
• Hampole et al. (2025) — 58M LinkedIn profiles, same cohort-specific pattern.
• Vicente & Matute (2023) — the Validation Tether's cleanest empirical anchor: 80.7% of participants detected the errors in biased AI recommendations and followed them anyway, then reproduced the bias in their own later judgments after the AI was removed. Surface validation worked; substantive validation failed. Detecting a problem and being able to override it are different capacities.
• Budzyń et al. (2025), Lancet Gastro & Hepatology — endoscopists' independent detection accuracy fell after adopting AI-assisted detection. The corpus's clearest measured deskilling case, in a safety-critical domain.
• Wiles et al. (2024) — AI assistance raised performance during access and produced no significant advantage on later unassisted assessment.
• Dell'Acqua et al. (2026) — elite consultants gained inside the AI capability frontier and lost outside it, unable to tell which side of the boundary a task sat on.
• Niederhoffer et al. (2025) — 40% of full-time employees received substantively flawed AI content in the past month, ~2 hours each to fix. Benzing et al. (2025) — 60% feel confident enough in AI output that they don't routinely check it.

The anchor, read firsthand (Canaries revision, August 2026)#

Until now the vault reached this page's load-bearing empirical claim only through Lovett's citation of it. It now holds the paper: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Brynjolfsson, Chandar & Chen, revised August 2026, ADP administrative payroll microdata through June 2026 (empirical; instrument caveats on Stanford Digital Economy Lab). Three things change.

The number, and the reason four numbers are circulating. The vault has cited 13%, 16% and now 19% as though they were a single widening series. They are not — the paper switched measures between vintages and says so:

Measure	July 2025 vintage	Sept 2025 vintage	June 2026 vintage
Regression adjusting for firm-level shocks (earlier headline)	13%	16%	de-emphasized
Simple descriptive "kept-pace" divergence (current headline)	15%	—	19%

"Earlier versions of this study headlined regression estimates adjusting for firm-level shocks (a 13% relative decline as of July 2025 data; 16% as of September 2025 data). We now emphasize the simpler descriptive divergence, which requires no modeling choices." So 13 → 16 is one series and 15 → 19 is another, and comparing across rows overstates the widening. In levels the current headline is arithmetic rather than estimation: employment of 22–25-year-olds in the two most-exposed quintiles fell ~11% between November 2022 and June 2026 while the same age group in the three least-exposed quintiles grew ~10% — a 21-point divergence, 19% relative to the less-exposed group's growth. About 60% of occupations in the least-exposed quintile saw rising early-career employment over the period against about 30% in the most exposed.

Mechanism 1 is now separated from displacement, and it is the shape this page predicts. Fact 4: the decline "operates primarily through reduced hiring rather than increased separations." Separation rates fell for both groups and, among young workers, fell at least as much in the most-exposed occupations as in the least — "the opposite of what displacement would imply." The gap is in hiring, opening after 2022. That is precisely this page's Mechanism 1 as distinguished from a layoff story: nobody is being removed from the ladder, the bottom rung is not being built. Fact 1 supplies the other half of the shape — no widespread, economy-wide job displacement (overall ADP employment +6% Nov 2022–Jun 2026, most-exposed quintile +4% pooling all ages). Healthy aggregate, failing regeneration, no bare hillside: the commons argument's signature, observed.

The mechanism has a first measurement. Codified vs Tacit Knowledge Exposure — young workers in high-codified occupations decline while experienced workers in high-tacit occupations grow faster — is the paper's own answer to "why the entry level," and it is the closest the corpus comes to measuring why the regeneration slot is the one being removed. Its honest asymmetry matters here: the codified (substitution) half does not survive a college-share control, the tacit (complement) half does.

Four counterweights, and one of them cuts at this page specifically.

1. Education attenuates it. The most-exposed-quintile coefficient for 22–25-year-olds is −0.179 with no controls and −0.091 with college share added (Table 1, Panels A and C; verified against pdftotext -layout), significant only at 10% in the latter. The authors argue education may be the channel rather than a confounder — AI substitutes for codified schooling — but concede the estimates "bracket a range."
2. Pre-trends. More- and less-exposed occupations diverged before ChatGPT, particularly around COVID. The paper's defence is a decomposition: the relative gap rose ~13 points to its mid-2020 peak and has since fallen ~30, of which ~13 unwind the boom and ~17 carry the gap below its 2018–19 baseline, continuing for three-plus years past the point where a reversion story predicts it stabilizes.
3. Magnitude is ADP-specific. Against the 2024 ACS the gap is −0.132 in ADP versus −0.022 [−0.055, +0.011] — direction consistent, magnitude not. The two agree closely inside professional/information/financial services (−0.231 vs −0.213).
4. The regression version of the headline got weaker while the descriptive version widened. Appendix L: under the August 2025 sample filters, the within-firm Poisson estimate for the most-exposed quintile at 22–25 was −11.7 log points (p=0.02); applying those same filters to the current data gives −5.3 log points (p=0.26) — no longer significant — while the fourth quintile is stable (−10.5 → −12.8, p=0.02). The authors attribute the sensitivity to data-pipeline changes and moved the headline to the descriptive measure partly because of it. Anyone citing this page's anchor should know that the estimate which conditions on firm shocks — the one that best rules out "these young people worked at firms that were shrinking anyway" — is the estimate that did not survive the revision.

What it does not touch. Nothing in 140 pages measures the developmental content of an entry-level role. Headcount, hiring rates, separations and base pay are all instrumented; whether the jobs that remain still teach anything is not, and could not be from payroll records. This page's central claim — that regeneration, not employment, is what is failing — remains exactly as unmeasured as it was.

What the framework says would falsify or bound it#

Unusually explicit boundary conditions, all of which cut against over-applying this page:

• Countervailing evidence acknowledged: collaborative AI workflows requiring active clinical engagement improved diagnostic accuracy (Everett et al. 2025); adaptive capacity in aggregate labor data (Manning & Aguirre 2026); heterogeneous retrainability (Hyman et al. 2025); null effects on earnings and hours in Denmark across the first two years of generative-AI adoption (Humlum & Vestergaard 2025). "These findings establish that the depletion mechanism is not universal."
• It excludes professions facing wholesale disintermediation (a different problem), and it assumes AI still needs human oversight — if AI reaches autonomous reliability, "the collective action problem shifts from expertise depletion to workforce displacement, a qualitatively different challenge."
• Unit of analysis is the occupation, not the economy. Nursing and the trades are low-vulnerability because novices learn through embodied practice AI doesn't substitute for.
• Five vulnerability factors: task substitutability, regulatory intensity, safety criticality, professional-association strength, work modularization. The predicted high-vulnerability set is software engineering, financial analysis, legal research — high substitutability, low regulatory intensity, highly modular. Medicine and engineering degrade slower because licensure and safety criticality create counter-pressure.
• Redistribution, not elimination. The paper's own hedge on its central mechanism: AI reallocates cognitive effort from generative struggle toward orchestration and evaluation, and "whether that reallocation builds or erodes expertise is conditional on how the work is designed." Assistance that preserves the practitioner's own attempt and makes them justify it can itself be developmental.

The time-delay problem#

Why nobody corrects course: "If organizations eliminate entry-level positions beginning in 2023, the labor market for experienced workers appears healthy because it reflects developmental investments from 2003–2020. The impact … will not manifest in experienced worker availability until 2030–2045." Stock depletion is measurable only through 10–20-year cohort analysis; functionality degradation — nominal seniors with shallower expertise than the previous generation — could show up sooner in validation error rates.

Paired with the Human Reserve Paradox: organizations need expertise held in reserve for crises and novel situations, but its value "remains latent until crisis reveals its absence. Like insurance, its worth becomes evident only when needed. Unlike insurance, no market mechanism exists to internalize commons maintenance costs across beneficiaries."

Governance: Ostrom, not Hardin#

The paper invokes Hardin "for its structure, not its fatalism," and hangs its prescriptive half on Ostrom's finding that commons are frequently sustained through local institutional experimentation — boundary definition, monitoring, graduated sanctions, collective choice, nesting. Three proposed levels, offered as hypotheses rather than recommendations:

• Organizational — phased AI introduction, minimum periods of independent human performance before AI-augmented work, AI-restricted deliberate-practice spaces. Grounded in Danry et al. (2023), where framing AI explanations as questions rather than answers improved logical discernment, and Everett et al. (2025), where requiring active engagement with AI reasoning preserved diagnostic performance. "How organizations design AI integration matters as much as whether they adopt it." The first randomized test of this rung is a learning study, not a workplace study: Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants (empirical, N=704; see Experimental Learning Impact of Generative AI). It keeps full AI access and only shows users what their own requests mean for their practice. That cut answer offloading (OR 0.47) and restored unaided performance to the no-AI level, while a points penalty on answer requests did neither. The data favour designing the interaction over pricing it.
• Professional association — polycentric experimentation with Ostrom's design principles; the framework predicts depletion bites hardest exactly where association capacity is weakest (technology, financial analysis).
• Public policy — investment rather than prohibition: training subsidies, credits for firms preserving pipelines, voluntary tiered credentials recognizing demonstrated Internalized Mastery, periodic AI-free competence assessment in continuing education.

Explicitly not the goal: resisting AI adoption or restoring pre-AI pathways. The stated target is keeping both masteries "in functional proportion."

Where this sits against the wiki's other evidence#

It contradicts The Automation–Optimism Link on mechanism, and the wiki should hold both. The AEI survey finds heavier delegators are more optimistic across all six job-quality dimensions, report their skills growing more valuable, and show no learning deficit. Lovett's Mechanism 2 predicts the opposite — that heavy delegation produces confident practitioners with degrading substantive validation. These are not directly comparable (self-reported sentiment vs a structural prediction), and the tiebreak already exists in the corpus: Experimental Learning Impact of Generative AI measures the thing both are guessing at, and finds the durable gains belong to augmentation users while automation users' gains vanish once AI is removed. That is Mechanism 2 observed under controlled conditions, in a one-week academic setting — which makes the open question whether it transfers to a career.

It supplies the missing theory for Unknowns as the Agentic Bottleneck's self-graded-quiz worry — the comfortable equilibrium where the quiz gets easier as the reviewer gets lazier is the Validation Tether degrading, and Vicente & Matute's 80.7% is the measurement of it.

It reframes Security Debt of Agent-Generated Code's central puzzle. That page finds review coverage of agent PRs converging toward the human baseline while catch-rate on leaked credentials sits near zero, with humans committing 67.6% of genuine credentials inside agent PRs. Surface validation present, substantive validation absent — the same shape, in a domain with a ground truth.

The vault's other framework paper, and where the two collide#

Banerjee & Singh's HAT model (arXiv 2607.20781, July 2026) is the corpus's only sibling to this page on method: a practitioner-opinion framework paper with no new data, deriving structural claims about AI and work from stated assumptions. Comparing how the two handle that shared position is more instructive than comparing their conclusions.

Scope. Lovett models the supply of validation-capable expertise and argues its regeneration mechanism is being removed. Banerjee & Singh model the demand side — the firm's risk-adjusted cost comparison that decides whether a given seat is filled by a human or an AI. Neither cites the other; they are the two halves of one question.

Method discipline is where they diverge, and this page comes out ahead. Lovett sorts his own claims into three explicit tiers (empirically grounded / theoretically derived / extrapolatory-untested) and states plainly that "the profession-level depletion this paper describes is a structural prediction … not an observed outcome" — the evidential ladder recorded above. HAT hedges globally rather than per-claim ("the results should be interpreted as conditional statements"), lists ten limitations at the end, and then publishes a seven-row table of "testable predictions" whose entries carry no per-prediction confidence at all. Its flagship result, middle-management vulnerability, is conditional on a single-crossing assumption whose feasibility curve is five numbers set by hand in the calibration. Same evidential position; one paper prices its own uncertainty into every claim and the other prices it into a preface.

The substantive collision is on HAT's P7. Its Theorem 15 models upskilling as a two-stage contest: the worker chooses adaptation effort u_j ∈ [0, ū_j] against a convex private cost, the AI side invests in capability and risk mitigation, and a Nash equilibrium determines who gets the task — with adaptation incentives strongest near the substitution boundary. That formulation assumes ū_j is a private endowment. This page's argument is that the developmental experience producing u_j is a profession-level commons whose regeneration mechanism — entry-level work — is exactly what AI removes, and that no individual firm has an incentive to maintain it. If that is right, ū_j is a shared, depleting ceiling that no agent in the contest controls, and the equilibrium HAT solves for is over a strategy set that is shrinking for reasons outside the game. The Human Reserve Paradox is the same point about R_ij: the value of retained human capacity is latent until a crisis reveals its absence, so a model that prices human risk from observable error, absenteeism and turnover rates systematically under-prices what the human was there for.

The practitioner denial, stated as strongly as it gets (DHH, August 2026)#

DHH holds the exact opposite of this page's structural prediction, in public, from the position of a 25-year open-source maintainer (Lex Fridman #501, 2026-08-26, practitioner-opinion). It is worth recording precisely because it is the strongest available statement of the position this page argues against, and because the two disagree on a factual claim rather than a values one.

His claim: expertise at the frontier does not accumulate at all.

"If I had just been backpacking for the last year, hadn't touched a computer, hadn't witnessed this agentic moment, and I just showed up yesterday — do you know what? I would've been caught up in two weeks. There's not any accumulation, which is in some ways a great relief… We have so many experiments running simultaneously right now that are constantly and ruthlessly sorting what works, what doesn't work. You don't have to remember or even be part of that entire journey. You can just show up for the results."

If true, the regeneration mechanism this page says is being removed does not need to exist for the agentic skill layer — there is no ladder because there is no accumulation. Two things bound the claim. It is scoped to tooling and frontier practice, not to domain judgment: the twelve months of harness churn he says you can skip really are mostly obsolete, which is not evidence about whether the ability to evaluate an agent's output can be acquired in two weeks. And it is a claim about himself — a person with forty years of computing behind him, for whom two weeks of catch-up is two weeks on top of the internalized mastery this page says newcomers will not be able to build. He never tests it against someone without that base.

His second claim: the commons gains, it does not deplete. He argues agents let people who could never contribute — non-programmers, programmers from other domains — land useful work, and that maintainers get to cherry-pick from a larger pool (Open Source Under Agent Contributions). This is orthogonal rather than contradictory: this page's concern is the supply of validators, and DHH's answer to who validates is that agents do, with a human making the final merge call — which is the Validation Tether restated as a solved problem without argument. Asked directly whether you still need the wizards to hold the bar, he says "100% you do," and never returns to where the next generation of wizards comes from.

Evidence weighting. This is one opinionated practitioner with a project to promote, against a framework paper that is itself unmeasured, with Experimental Learning Impact of Generative AI as the only controlled measurement in the neighbourhood and it favors this page. Nothing here is superseded; the disagreement is logged, and the honest state of the question is that both sides are arguing from structure and neither from a longitudinal measurement of practitioners.

The practitioner affirmation, from the education side (Ng, August 2026)#

Two days after DHH's denial, Andrew Ng — whose career is AI education — states Mechanism 2 as flatly as Lovett does, from the opposite chair (Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think, Silicon Valley Girl, 2026-08-28, practitioner-opinion). "When you ask AI to do work for you you're cognitive offloading to AI which is great because that's how society moves forward and gets work done but human retention is much worse." His illustration is Internalized versus Distributed Mastery in one person: six months after having a model explain a front-end/back-end component, "I don't remember the answer… I ask AI again." He worries about it most for children — he withholds the calculator from his five- and seven-year-old "and practice that with them," and "as they're a little bit older, I worry a lot about students using cognitive offloading to AI in a way that damages the long-term learning retention." This is the framework's premise, endorsed by someone with the opposite commercial incentive to DHH: he sells learning, and was announcing a new venture, LearnVector, built to fix exactly this.

Two things bound the affirmation, and they pull the other way from DHH's. First, the regeneration slot in his own account is open: his office is full of "interns… current college students, fresh college grad… one high school intern" who are "amazing and productive," and employers he knows "can't find enough skilled… people at any level of seniority" — the entry-level work exists, at his firm, for the AI-native. What he says has failed is the institution that used to feed it: universities "take like a year or two" to teach a new skill and are "still teaching students to be ready for the jobs of 2022," so the pipeline has rerouted through online courses — his own. That is not depletion of the commons but a claim about who now maintains it, and the maintainer he names is a private education company. Second, he scopes the retention claim to "LLMs as they are most commonly used" and concedes there are ways to use them that teach — the augmentation half of Experimental Learning Impact of Generative AI, the same controlled measurement that already sits between this page and DHH's denial. Nothing here is measured either; it is one more practitioner arguing from structure, on the framework's side.

The same mechanism, reached from the delegation boundary — and the argument that it is privately rational (CIVIC-AI, September 2026)#

When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration (CIVIC-AI 2026 workshop whitepaper, 22 authors across NUS, Stanford, A*STAR, NTU and Singapore's Ministry of Manpower, practitioner-opinion, no data of its own — its framework is carried on Human-AI Accountability Redesign) makes two of its six conditions for "genuine augmentation" deepening learning and career pathways, and arrives at this page's mechanism from the opposite end. Two framework papers with no measurements agreeing is not evidence, and it is not counted here as any. What is worth keeping is the route, which is different enough to sharpen the claim, and the prescription, which contradicts this page's.

The route: delegability and pedagogy are the same property, not two properties that happen to overlap. Lovett's argument runs from the commons — entry-level work regenerates expertise, AI removes entry-level work. CIVIC-AI runs from the delegation criterion, and lands on the collision as a selection effect: "the tasks most amenable to AI delegation tend to be high-volume, procedurally defined and verifiable. These are exactly those tasks through which junior workers develop their tacit judgement needed to evaluate outputs, catch failures and push back on algorithmic recommendations." The whitepaper's own boundary rule (delegate where outputs are inspectable, errors reversible and stakes bounded) therefore selects on the same variable the developmental pipeline selects on. That is a stronger statement than "AI happens to take the junior work": a rational deployer following the safest possible delegation rule strips the training ground first, and would do so even with no cost pressure at all. It also predicts where the depletion will not happen — wherever the delegable-and-formative sets come apart, i.e. work whose formative content is unverifiable or high-stakes from the start.

The prescription, and where it cuts against this page. Lovett's conclusion is Ostrom: a profession-level commons needs profession-level governance, because the firm that maintains regeneration bears the cost and shares the benefit. CIVIC-AI proposes a firm-level staffing rule instead — reviewer competence "does not persist automatically. It requires organisations to deliberately assign workers enough substantive review work and enough exposure to AI failure to keep the verification skill current" — and defends it on self-interest rather than equity: "preserving learning and career pathways is not merely an equity concern, but rather a pre-requisite for ensuring human control stays meaningful over time." In its condition ordering, Layer 2 (learning, pathways, purpose) is upstream of Condition 2 (meaningful human control), which is upstream of the firm's own accountability and liability. If that ordering holds, part of the externality is internalized: a firm that stops regenerating validation capability loses its own oversight, on its own clock, not just the profession's. Its worked example gives the lever a concrete form — junior researchers must perform initial protocol design and coding themselves, justified by "anchoring bias and potential for vacuous verification of an AI's generation," which is Layered Supervision's human layer specified as an input rather than a posture.

The honest weighting. This is the Ostrom claim's first named rival mechanism, not a refutation of it, and the two are separable by an observation neither paper has: whether the firms that do protect formative work are protecting it on the timescale of their own oversight needs (the CIVIC-AI story) or under external pressure from regulators, professional bodies or unions (the Ostrom story). Note also that the internalization argument only works where the firm's own reviewers are the depleted population; it says nothing about the case this page is actually about, where the depleted population is the next generation of the profession and the firm's current reviewers were trained before adoption. The time-delay problem above is untouched: a firm's oversight degrades slowly enough that the private incentive arrives late, which is why the whitepaper's own recommendation ends in a public instrument — a shared workflow record, sector thresholds, and labour-ministry measurement — rather than in the staffing rule alone.

Connections#

• Seniority-Biased AI Adoption: The Junior Share at Adopting Firms — the first firm-level, 41-country evidence that AI adoption shrinks the junior share (−1.9pp), mostly by growing seniors, with the steepest within-occupation declines in legal, computing and business/finance

• The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay — the price side of the entry-level event: entry offers in exposed occupations got no premium, and the headline premium is mostly the entry-to-senior posting shift

• AI Adoption in Scientific Work — this argument reached independently from philosophy of science and applied to research: replacing doctoral and postdoctoral work removes the channel that transmits tacit knowledge, methodological judgement and professional norms, producing rising scientific output alongside falling opportunities to become a scientist. No measurement behind it, and it insists on the symmetric branch — integration that supports researchers could raise scientific employment instead

• The Enablement–Regulation Axis — the public-policy rung of the Ostrom ladder, priced against what legislatures actually say. Lovett's third level asks for training subsidies, credits for firms preserving entry-level pipelines, tiered Internalized-Mastery credentials and periodic AI-free competence assessment — investment, explicitly not prohibition. Chueri & Törnberg's 33-parliament corpus finds the training frame at 20.6% of response-frame mentions and invoked at broadly similar rates across every party family, precisely because it is politically portable: it serves adoption, adaptation and national skills capacity at once. So the rung this framework reaches for is the cheapest one on the board and carries no positional information. The rung the argument actually needs is a different one — a credit for preserving junior work is a constraint on labor-replacing adoption, which lives in the 21.8% regulation-and-restriction frame, is concentrated in the radical left and greens, and inside that frame is outweighed nearly three to one by copyright and creative-sector protection (36.1% vs 27.2% for displacement safeguards). The commons has no parliamentary constituency yet; its adjacent vocabulary has a very popular one

• Layered Supervision — the human layer this page's Validation Tether is the missing input to: supervision re-scoped to concurrent oversight and "operational explainability" presumes a reviewer whose judgement is still being regenerated, and the paper measures neither the layer nor the regeneration

• The Solo-Founder Shift — the same left-tail shift at firm scale, and a case where the apprenticeship reading weakens: a co-founder slot that never formed is not obviously a training slot destroyed, since solo founders hire their first employee sooner than co-founded teams do

• The Solo-Authorship Rebound — the mechanism's precondition measured, in one profession and at population scale: across 300M+ OpenAlex works the decades-long decline in solo authorship halts at ChatGPT's release, strongest in fields where a coauthor's work is most substitutable, and it survives conditioning on authors who had never published alone. Matsui's own Implication 2 is this page's argument verbatim — "the more senior researchers can hand execution work to LLMs instead of to junior coauthors, the weaker this training pathway through collaboration may become." What it establishes is the removal of the apprenticeship slot, not the depletion of the commons: the seniority gradient carrying that reading is weak and filter-dependent, solo papers stay a small minority, and nothing there measures whether any junior learned less

• Owning Your Externalized Cognition — the same asset's other failure mode, and it aggravates this one. Where this page describes expertise that never forms because the entry-level work is gone, Tan describes expertise that forms normally, gets written down as executable skill files, and then accrues to whoever holds the repo. Internalized Mastery built through years of cognitive struggle is exactly what makes a skill library worth extracting. practitioner-opinion with no measurement, and neither author has read the other

• Codified vs Tacit Knowledge Exposure — the mechanism under this page's premise, measured for the first time: AI substitutes for the codified knowledge that entry-level work used to be paid for and complements the tacit knowledge only experience produces. If both halves hold, the slot being removed is specifically the one through which tacit knowledge was acquired — which is this page's argument with an instrument attached. The codified half does not survive an education control; the tacit half does

• Returns to Expertise in Agentic Coding — the direct tension: expertise measurably amplifies an agent today, and this page argues the pipeline producing that expertise is being cut. Both can hold — the current expert stock is intact, the regeneration is what's in question

• Firm AI-Spend Intensity and Headcount Growth — the two labor-demand instruments that bear on the regeneration question, pulling opposite ways. Indeed's job-postings rebound in the most-exposed occupation is 71% senior, the seniority-biased shape this page predicts; Ramp's spend-linked firm panel finds entry-level headcount growing fastest (+12.0%) at intensive adopters, which is the pipeline reopening at exactly the firms doing the adopting. Different units (vacancy flow vs. headcount stock) and different populations, so neither refutes the other — and neither measures the developmental content of the roles that came back

• Experimental Learning Impact of Generative AI — the closest thing to Mechanism 2 measured: augmentation users keep the gains, automation users lose them once AI is removed. Controlled, one week, students

• The Automation–Optimism Link — the contradicting finding on sentiment and self-assessed skill growth; different instrument (self-report), different claim (felt vs structural), and the wiki keeps both

• Unknowns as the Agentic Bottleneck — the practitioner-side statement of the Validation Tether: reviewing your own delegated work with a quiz you also wrote

• Security Debt of Agent-Generated Code — surface-vs-substantive validation with an outcome measure attached: review happens, credentials still ship

• Review as the Control Point — the control-point argument depends on reviewers who can substantively validate; this names what erodes that capacity

• Task Crossover — the empirical pattern this frames as a risk: work moving to whoever encounters the need is exactly work moving away from the specialist who would have built mastery on it

• AI Brain Fry — the other cost of oversight: brain fry measures the fatigue of validating, this measures the erosion of the ability to validate

• Jagged Intelligence (Ghosts, Not Animals) — Dell'Acqua's consultants failing to locate the capability frontier is the jagged-edge problem restated as a validation requirement

• Outsource Your Thinking, Not Your Understanding — the individual-level version of the same prescription, here scaled to a profession and given a collective-action reason it won't happen voluntarily

• Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — the five vulnerability factors are an exposure measure on a different axis: not what share of tasks AI can do, but whether the tasks it takes are the ones novices learn on

• Organizational Complements to AI — developmental infrastructure as a complement nobody is incentivized to supply, because its benefits are non-excludable. Also the home of the HAT substitution model, this page's sibling in method (no-new-data framework paper) and its collision partner on P7: upskilling as a private effort choice against a commons the individual does not control

• Human-AI Accountability Redesign — the org-design prescriptions overlap; this adds the level above (profession and policy) and the reason org-level action alone is insufficient

• Verification as the New Bottleneck — the Validation Tether is the bottleneck's precondition: verification capacity is not just scarce time, it is a depreciating skill

• The Data Wall and the Validation Commons Are One Supply Constraint — the ordering risk named in the next bullet, searched for across the corpus and answered. No domain shows commons-scale depletion (this page's own disclaimer is the strongest evidence for that), the closest measured instance is Budzyń's colonoscopy deskilling — which satisfies erosion-before-threshold but is atrophy of the existing stock rather than regeneration failure, and sits in this framework's low-vulnerability set. The correction it offers: the threshold arrives first exactly where a sound cheap verifier exists (Lean, hidden test cases, execution feedback), which is where validators were never load-bearing, so arrival order and need order are correlated rather than independent. The residual risk is real one level down, at the sub-task boundary. It also finds this page's commons argument restated verbatim inside the training literature — CS329A's motivation for self-proposed curricula is that the supply of experts who can write the next question runs out as models pass human expert level

• Post-Scarcity Macroeconomics — the direct denial of this page's premise: Musk's "Stockfish level" threshold is the claim that validation stops being needed, which would make the commons' depletion costless rather than catastrophic. The two are incompatible and neither offers an advance test — but note the ordering risk, since a commons destroyed unevenly is destroyed before the threshold arrives in the domains that still need validators

• Outsource Your Thinking, Not Your Understanding — the individual-scale version of the same erosion, and the first proposed instruments for it: Thawar's two-to-three-layers-down requirement and weekly demos as a comprehension signal the Validation Tether otherwise has no way to check

• Open Source Under Agent Contributions — DHH's counter-position from the maintainer's chair: the contribution pool grows, triage delegates to agents, and the Validation Tether is treated as solved by delegating validation itself

• DHH (David Heinemeier Hansson) — the practitioner holding the strongest available denial of this page's prediction: "there's not any accumulation"

• Andrew Ng — the practitioner affirmation of Mechanism 2 from the education side, with the regeneration slot reported open at his own firm and the university named as the failed institution

• Does the Augmentation/Automation Split Govern Skill at Work? — this page's borrowed evidence sorted onto the augmentation/automation rows of Experimental Learning Impact of Generative AI: Budzyń, Dell'Acqua, Wiles and Vicente & Matute all show the removal signature (performance high with the tool, no advantage without it), Everett shows the augmentation row, and the separating variable is the one this page already names — whether the practitioner's own attempt is preserved. It also carries the tier discount: this page is the vault's only route to all six, so the workplace pattern is an assembled bibliography rather than a measurement

• Psychological Costs of AI Adoption — Mechanism 2 heard in the first person, which is not the same as measured. Practitioners at a regulated-software firm one year into AI adoption name this page's erosion unprompted and in its own terms — "if we lose access to these things... will I still be good at it?", "if you use it too much, you will forget some of the skills", and a test manager anticipating "skill atrophy" that might eventually leave engineers "unable even to critically review AI-generated solutions," which is the Validation Tether stated by someone who has never read the paper. What follows cuts both ways for this page. The defence is voluntary and individual: daily manual-coding rituals, drafting a solution before consulting the model, keeping business-logic work and delegating only boilerplate. That is better news than a commons model predicts — the resource is being conserved by its users without governance — and worse news than it looks, because none of it is organizationally supported, budgeted or even recognized, so it survives only as long as practitioners retain the discretion to spend time that way. The cohort is also the wrong one for this page's core claim: median 20 years of industry experience, 19 of 21 with a decade or more, so their validation capability was accumulated before the entry-level work disappeared. What they are defending is a stock, not the regeneration mechanism

• Pilot-to-Production Gap — the prescription this page's argument contradicts, written by parties who do not cite it. Anthropic × Accenture's consideration 05 prescribes converting operators into overseers as AI adoption spreads — work shifts "from executing tasks to overseeing a system", requiring judgment-based skills (spotting incorrect outputs, deciding what escalates, detecting drift), treated as a training-and-role-design investment. Lovett's mechanism says the skill being invested in is regenerated by the entry-level execution work the same programme removes, so an organization that converts everyone to overseers has automated away the training path for the capability it now depends on. The document's worked example is the collision in miniature: "the claims processor who used to review every invoice now audits the system that reviews them." vendor-claim prescription against this page's practitioner-opinion argument — neither side is measured, and the falsifiable version is filed as an open question on that page

Open Questions#

• Mechanism 2 has never been directly measured in a workplace. Does a cohort that entered an AI-heavy profession after 2023 show lower unaided task accuracy than a matched earlier cohort at the same tenure? The paper specifies the design (no-AI assessment stratified by cohort and AI exposure); nobody has run it. Related (2026-09-05): Does the Augmentation/Automation Split Govern Skill at Work? searched the vault for a substitute and confirms there is none — the four workplace results with the removal signature (Budzyń, Dell'Acqua, Wiles, Vicente & Matute) are all borrowed through this page, and the vault's own workplace instruments measure delegation, throughput, error rates and retention but never unaided skill. It also names a second design that would bear on the same gap: classify use mode within-subject from usage logs, then measure performance after AI is withdrawn. Related (2026-09-22), not an answer: The Psychological Costs of Artificial Intelligence Adoption in Software Engineering (case-study) is the corpus's first workplace source in which practitioners state the mechanism themselves — anticipated skill atrophy, and an explicit worry about eventually being unable to critically review AI output — and the first to inventory the counter-practices they adopt against it. It measures no unaided accuracy of any kind, its cohort is 19-of-21 veterans whose capability predates the exposure, and self-reported fear of deskilling is exactly the self-report this question was written to route around. It shifts the mechanism from inferred to voiced, which is worth one line and nothing more.
• The cohort evidence is a snapshot ending Sept 2025 in the most AI-exposed occupations. Does the 22–25 employment decline persist, reverse, or re-sort as agentic tooling matures — and if entry-level postings recover, does that restore the developmental content of the work or just its headcount? Recovery of positions and recovery of the regeneration mechanism are not the same event, and only the first is currently instrumented. Partially answered (2026-08-04): Indeed Hiring Lab instruments the first half — postings in the most-exposed occupation did rebound (US software development +15% since February 2025 against overall postings −7%) — and the composition answers the sub-question in this page's favour: 71% of the May 2025 – May 2026 increase is senior roles, 37% AI-titled, with the author himself conceding the market "could still be experiencing a seniority-biased technological change." So the recovery is real and is not entry-level, on this instrument. It leaves the harder half untouched: nothing there measures the developmental content of any role, the data are one job board's vacancy flow analyzed by that job board, and Ramp's firm panel finds entry-level headcount growing fastest on a different unit. Extended (2026-09-22) — the first clause is now settled and the second is not, which is why this stays open. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (empirical, ADP payroll microdata through June 2026) answers "persist, reverse, or re-sort" without ambiguity: it persists and widens, from a 15% kept-pace shortfall at the July 2025 vintage to 19% at June 2026, with the divergence continuing for more than three years and through a period in which interest rates stabilized and fell. It also re-sorts in a specific way rather than diffusely — the loss concentrates where AI usage is automative (−0.098 per SD at 22–25, shrinking monotonically to −0.006 at 50+) and is absent where it is augmentative, and it runs through hiring rather than separations, which is this page's Mechanism 1 stated as a measurement. The snapshot-ending-Sept-2025 premise in the question above is therefore obsolete; treat the first clause as answered. The second clause is untouched and unreachable from this instrument: payroll records carry headcount, hiring rates, separations and base pay, and carry nothing about whether the roles that remain still teach. Until something measures developmental content, the question this page actually cares about has no evidence on either side. #oq/source Extended (2026-10-01): Indeed corroborates persistence through 2026 on postings — entry share in the most-exposed occupations 29% → 10% since 2021; see The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay. Extended (2026-10-01), to firms and outside the US: Chandar & Klein Teeselink (empirical, Revelio, instrumented) finds the junior share at AI-adopting affiliates −1.9pp against matched non-adopters by March 2026. The share is negative in 23 of 31 countries and significant in seven (US, UK, Spain, Poland, Brazil, Mexico, Saudi Arabia). It comes mainly from senior growth (+6.7%) rather than junior cuts (−2.5%, n.s.). So on this instrument the regeneration slot is shrinking as a share at adopters, not necessarily as a count. The developmental-content clause is still untouched; see Seniority-Biased AI Adoption: The Junior Share at Adopting Firms. Extended (2026-10-01), to graduates by major: Census PSEO×LEHD records show the most-exposed decile of majors still −2.0pp employed and −3.3 log points in earnings two years after graduating, while deciles 7–9 recover. The move into retail and food service is the nearest proxy yet for first-job content, but it is a sector code, not a measure of learning; see AI-Exposed College Majors at Labor-Market Entry. Counter-evidence (2026-10-01), reconciled rather than superseding: Fairlie & Wu (CPS, empirical) find no significant summer-2026 unemployment break for all 22–25 bachelor's holders, including a 'sidelined' measure. It measures a different outcome (unemployment, not occupation headcount or earnings) in a wider population, so the persistence clause stands; see AI-Exposed College Majors at Labor-Market Entry. Related (2026-09-22), and deliberately not counted as partial: When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration draws this bullet's distinction independently — entry-level PMET openings in Singapore rose 32,500 (Dec 2025) → 32,800 (Mar 2026) with fresh-graduate outcomes "broadly resilient," which its authors read as snapshots that "do not yet indicate broad-based deterioration in entry-level opportunities, but neither do they establish the longer-term effects of AI on human development." Positions and regeneration are separated there for the same reason they are separated here, and the instrument proposed to close the gap is the one this bullet says is missing: repeated longitudinal measurement of learning, progression and agency inside the workflow, not headcount at the door. It answers nothing — the figures are third-party MOM statistics reaching the vault secondhand in a practitioner-opinion framework paper, a single small economy, one quarter apart, and no developmental content is measured in either jurisdiction.
• The framework predicts differential depletion by its five factors. Do software engineering, financial analysis and legal research actually diverge from medicine and engineering on validation-capability measures — or does regulatory intensity turn out to be weaker protection than the model assumes? Related (2026-09-22), and deliberately not counted as partial: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence publishes occupation-level detail at last — the most-exposed quintile's largest ADP occupations are chief executives, customer service representatives, accountants, HR specialists, wholesale sales reps, systems software developers and computer/IS managers, with bookkeeping and billing clerks further down (Online Appendix Table A.6, ranks 1–25 verified; the table's tail is parse-damaged and is not cited). That overlaps two of the framework's three predicted high-vulnerability professions and misses the third entirely — legal research does not appear, while clerical and customer-facing work the framework does not theorize about dominates the list. More to the point, the paper never cuts employment change by profession, licensure or regulatory intensity, so it supplies an exposure ranking and no test of the five factors. What it does supply is a warning about the question's framing: the five-factor model predicts depletion in high-skill, low-regulation professions, and the measured decline is concentrated in occupations that are exposed rather than professional. Partially answered (2026-10-01), on the input side only: Chandar & Klein Teeselink cut the within-occupation junior share at AI adopters by profession. The significant declines are legal −4.9pp, computer & mathematical −3.3, business & financial operations −3.2, plus sales and management. Healthcare practitioners and architecture & engineering have intervals spanning zero (chart reading). That is the framework's predicted high/low split, profession by profession, on entry-level headcount. It is not validation capability, legal does not survive Bonferroni, and the country-level labor-rigidity correlation (+0.20, n.s.) measures general labor law, not licensure.

Sources#

• How Does AI Change Labor Demand? Evidence from 41 Countries — Chandar & Klein Teeselink, Stanford Digital Economy Lab working paper, 2026-09-20, 156pp (empirical). Cited only for §1/§5.1 headline composition numbers, §5.1 Figure 5's within-occupation junior-share ordering, and §5.2.1's country count. Full treatment on Seniority-Biased AI Adoption: The Junior Share at Adopting Firms

• Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants: Maier et al., arXiv 2609.20143 (2026-09-17, empirical). Cited only for the organizational-rung sentence under Governance. Full treatment and parse notes are on Experimental Learning Impact of Generative AI.

• DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501 — DHH, Lex Fridman #501 (2026-08-26, practitioner-opinion): the no-accumulation claim ("caught up in two weeks"), the growing-commons argument, and the unexamined concession that the wizards are still needed

• The Human-AI Substitution Principle: When will you be replaced by AI in your organization? — Banerjee & Singh, arXiv 2607.20781 (2026-07-22; practitioner-opinion, formal model, no empirical data): §4.5.3 Theorem 15 (the two-stage upskilling-vs-AI-investment contest and its private effort set), §5.6 Table 3 prediction P7 + §5.6.7, §4.3 Corollary 2 and §5.7 (the hand-set F(i) behind middle-management vulnerability), §1 and §5.8 (its global conditional hedge and ten limitations). Cited here only for the method and P7 comparison; full treatment and the P1–P7 ledger at Organizational Complements to AI

• AI and Job Postings: From Destruction to Creation? — Guillermo Gallacher, AI and Job Postings: From Destruction to Creation? (Indeed Hiring Lab, 2026-07-08; empirical postings data, blog post, analyzed by the platform that owns the data). Cited here only for the postings-recovery question: the key points list (+15% software development / −7% overall since February 2025) and §"A senior, AI-fluent rebound" (71% senior, 37% AI-titled) plus the conclusion's seniority-biased-technological-change concession. The post's causal framing — that agentic coding tools drove the rebound — is a coincidence of timing with no control group; see the evidence note at Firm AI-Spend Intensity and Headcount Growth.

• The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise — Nolan Lovett, The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise, arXiv 2607.29380 (v1 2026-07-31); author accepted manuscript of the article in Human Resource Development Review (Sage), advance online publication 2026-07-26, doi:10.1177/15344843261470602. practitioner-opinion — peer-reviewed but conceptual; contains no original measurement. Parse note: docling verify: ok on all checks; the 6 tables are definition/indicator grids with no numeric cells, but rows wrap across several markdown rows with blank leading cells (Tables 1 and 2, raw lines ~214–282) — read them by stitching consecutive rows. All quantitative claims on this page are quoted from the paper's prose, not its tables, and are attributed to the original studies. Canary-recall reconciliation (2026-08-04): body prose verified intact (15/17 after splatter folds); the two remaining misses are References-section page-ranges docling truncated to a welded DOI (78-98. → 78 -https://doi…) — the raw's bibliography page numbers are unreliable, article content unaffected.

• Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think — Andrew Ng interviewed by Marina Mogilko, Silicon Valley Girl (2026-08-28, practitioner-opinion): the cognitive-offloading claim, the calculator-and-kids passage, the AI-native interns, and the university-lag argument. Two days after the DHH episode; neither refers to the other

• Deploying AI from pilot to production: A practical blueprint for CIOs and technical leaders — Deploying AI from pilot to production, Anthropic × Accenture, 2026-09-11, 38pp, vendor-claim. Cited here only for the consideration-05 oversight-role prescription, as the counterposition to this page's mechanism. Full treatment and evidence limits on Pilot-to-Production Gap

• AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It — Kennedy, Indeed Hiring Lab, 2026-09-17 (empirical); entry-share figure only, see The AI-Exposure Pay Premium: Advertised Offers vs Realized Pay

• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — Brynjolfsson, Chandar & Chen, Canaries in the Coal Mine?, Stanford Digital Economy Lab, revised August 2026, 140pp (empirical, ADP payroll microdata through June 2026). This page's anchor, now read firsthand instead of through Lovett's citation. Cited for the abstract and §1 (the 19% kept-pace shortfall and the vintage ladder), §2.1–2.2 (no economy-wide displacement; the ~11% / ~10% quintile levels; Table 1's −0.179 / −0.091), §2.3 (the widening), §2.4 (hiring rather than separations), §2.5 (automation vs complementarity, Table 3), §4.1–4.2 (education attenuation, pre-trend decomposition), §5 (ADP vs ACS), Online Appendix Table A.6 (quintile-5 occupations, ranks 1–25 only) and Appendix L (the vintage sensitivity of the within-firm estimate). Parse notes: table-collapse fires 15 times on repeated panel-header cells — a benign row-span artifact, confirmed against pdftotext -layout, with Tables 1 and 3 reconciling cell-for-cell. Table 2 is genuinely row-shifted (values slide sideways into adjacent columns) and is cited nowhere. Table A.6 is split-row/welded past rank 25; only ranks 1–25 are used. Full ledger in Source Notes

• When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration — Wu, Ziems et al. (22 authors incl. Singapore Ministry of Manpower), When Does AI Augment Work?, CIVIC-AI 2026 workshop whitepaper, arXiv 2609.12482, 2026-09-11, 8pp, practitioner-opinion, no original measurement (§5's worked example is a proposed pilot; §6 grades third-party statistics). Cited here for §4 Layer 2 (the delegable-equals-formative selection argument, the competence-maintenance staffing rule, and learning/pathways as a prerequisite for control rather than an equity concern), §5 (the anchoring / vacuous-verification rationale for juniors doing initial coding themselves) and §6 (the Singapore entry-level PMET and fresh-graduate figures, secondhand from MOM and not independently verified). Full framework treatment at Human-AI Accountability Redesign