# ATTACK NOTES — *The Third Disembedding*, draft 2.0 Written in the voice of the most hostile competent referee the paper is likely to get: a labour economist who finds Polanyi embarrassing, reads the model as a costume, and has the empirical literature at their fingertips. Nothing here is softened. The point of the file is that the paper should have to survive it before anyone sees it. Filed 2026-07-31. Updated for draft 2.0, which adds Marx as a second pillar and a limit-case section, and applies 24 corrections from a second independent verification pass — including one place where the *first* pass corrected a figure in the wrong direction. --- ## The kill shot > *A framework with a negative result, dressed as a discovery. The author measured the > thing his thesis needed, found nothing, invented an unmeasurable substitute called > "expertise rent," and then declared that the substitute is the real fictitious > commodity. The measurable version failed; the unmeasurable version cannot. That is > not rescuing Polanyi from unfalsifiability — it is reproducing it one level down, > with a task-based model bolted on for respectability.* That is the sentence to beat. The paper's defence — that the substitute carries a stated prediction with an independent shadow measure — is real but is currently carried by **one** field experiment in **one** occupation at **one** firm. **Status: not yet unfair.** It becomes unfair the day Prediction 1 has been run. --- ## 1. The empirical core is eleven observations and a vendor's telemetry This is the worst problem and the paper knows it, which helps less than the author thinks. - **n = 11.** Both headline correlations have bootstrap intervals spanning nearly the whole feasible range: [−0.62, 0.59] and [−0.66, 0.29]. Neither is significant at any conventional level. The paper's central empirical claim — the sign of *c* — is not established. It is *suggested*, by a point estimate that a single reclassified occupation could flip. - **The AI measure is not an exposure measure.** It is the share of one vendor's chatbot conversations divided by employment share. Anthropic states the skew: computer and mathematical occupations are ~30% of respondents against 4% of employment. Usage is endogenous to exactly the thing being studied. Several values were read off a figure. - **Weighted versus unweighted.** Section 7 defines *c* as a wage-bill-weighted correlation across a task continuum. Section 9 computes an unweighted correlation across eleven occupation groups and feeds it into a simulation calibrated on tasks. The paper now flags this. Flagging is not fixing. - **Occupation groups are not tasks.** The whole model is task-level; the whole measurement is occupation-group-level. Within-group heterogeneity in both exposure and protection is enormous and is being averaged into nothing. **Verdict: the paper's negative result is the honest reporting of a null, and it is being asked to carry more weight than a null can carry.** The correct statement is "there is no detectable relationship," not "the relationship is negative." **Fix:** Study B from §12.2, done properly, before any further claim about *c*. --- ## 2. The historical calibration has two dating problems and the paper only owns one - **T2 is dated to 1929.** That is a demand collapse, not a wage–productivity divergence, so it violates the paper's own stated dating convention. The New Deal responded to the Depression, not to electrification. The paper now concedes this, which is correct and also devastating: **the six-year figure is the number that makes T3's forty-six years look catastrophic.** Take T2 out and the "lag is not falling" claim rests on two observations, one of which has no endpoint. - **T1's response is mis-specified too.** The Factory Act of 1833 regulated child labour in textile mills. It was not a response to hand-loom weaver displacement. Conceded in the current draft, but the concession is a paragraph and the three-observation table is a figure. - **Three observations, contestable dating, no endpoint on the third.** This is not a calibration; it is three anecdotes in a table with Greek letters attached. **The honest residual:** T3 is genuinely informative on its own. Forty-six years of measured harm with contracting collective-bargaining capacity is a fact, and it does not need T1 or T2 to be an embarrassment for Polanyi's functionalism. The paper would be *stronger* if it dropped the τ series and argued only from T3. --- ## 3. "Expertise rent" is doing enormous work on very little evidence The substantive contribution rests on a chain: 1. Professional occupations have wage premia not explained by union or licensure → 2. the residual is a scarcity rent on tacit knowledge → 3. that rent was sustained by Polanyi's paradox → 4. ML dissolves the paradox → 5. therefore expertise is being commodified in Karl Polanyi's sense. Step 2 is a residual argument — "what else could it be?" — and the honest answer is: selection, unmeasured ability, compensating differentials, firm-specific capital, credential signalling, rent-sharing in high-markup industries, and simple skill-biased technical change. The paper does not rule out any of them. **A residual is not a mechanism.** Step 4 is asserted from Brynjolfsson–Mitchell and Kambhampati, both of whom are making claims about *capability*, not about *rent extinction*. Nobody has shown the rent falling. **The one real piece of evidence** — Brynjolfsson, Li and Raymond's +34% for novices against ~0 for the experienced — is a productivity result in a call centre. Extending it to "the professional wage premium is a tacit-knowledge rent that AI is extinguishing" is a leap of several orders of generality. **Fix:** run Study A. Until then, §10 is a conjecture and should be labelled as the paper's *hypothesis* rather than its *finding*. The current draft comes close to this but §10's prose is more confident than §12's table. --- ## 4. The model is not doing the work the paper claims - **Proposition 1 is an identity.** dD/dI = (ω/W)(m(I) − M(I)) is just the derivative of a conditional mean with respect to a truncation point. It is true of any variable and any ordering. Calling it "the whole Polanyian mechanism in one line" dresses a statistical triviality as a theoretical result. It *is* useful — it tells you what to measure — but it is not a theorem about the economy. - **The production block is decorative.** σ, the CES structure, the weak-links wage — none of it enters Proposition 1, which depends only on the protection distribution and the ordering. The paper could delete §6.1 and lose nothing except the appearance of being a macro paper. The wage index appears in the simulation only as a harm generator. - **The countermovement response function is unidentified.** λ and τ and the harm threshold are three free parameters generating any trajectory you like. The "race condition" ρ* = 0.29 is a property of the simulation's parameterisation, not of anything in the world. - **∂²H/∂g∂τ > 0 is not derived.** The first-order expression printed in §11.1 contains no *g*. The cross-partial comes out of the simulation. The paper now says so, which converts a stated theorem into a numerical observation about a toy. **Verdict: the formal apparatus is a clarifying device, not a source of results.** That is a legitimate use, and the paper should say so plainly rather than presenting propositions. --- ## 5. The Polanyi scholarship is careful but the framework may still not survive it - **Gemici (2015) is the objection the paper answers least well.** If the self-regulating market just *is* the neoclassical model, the disembedding gradient measures distance from a fiction. The reply — that the index measures observed institutional coverage — is decent, but then the index is a labour-institutions measure and the Polanyi framing is ornamental. **Ask: what does invoking Polanyi buy that "union coverage fell and professional rents are eroding" does not?** - **Dale's objection is conceded without an answer.** The paper admits that using Polanyi to license regulated capitalism is a misuse of Polanyi, says the machinery is separable from the politics, and moves on. A referee sympathetic to Polanyi will read that as extraction. - **The Speenhamland concession is excellent and also double-edged.** Having established that Polanyi's leading historical case is empirically indefensible and that he was told and published anyway, the paper then asks the reader to take his mechanism seriously. The obvious retort: *why?* --- ## 6. Things a referee will simply catch - The paper is a working paper by an independent author with no identification strategy, no standard errors on anything except a bootstrap over eleven points, and no causal claim. It will be desk-rejected at any journal that wants an estimate. Its natural home is a discussion paper series or a heterodox journal, and the latter will object to §4 and the former to §2. - **JEL codes claim more than the paper delivers.** J51 (labour–management relations) is claimed on the strength of a union-density chart. - The disembedding index has no external validation. Nobody has ever measured it. There is no series to compare against, no benchmark, no test of whether it tracks anything. - **The abstract still oversells.** "Expertise is thus a fictitious commodity in Karl Polanyi's definitional sense" is stated as a conclusion; §9 lists it as a conjecture with a test that has not been run. --- ## 6b. New attack surface in draft 2.0 **The Marx sections invite three specific hits.** 1. **"You bracketed the value theory and kept the conclusions."** The paper takes the reserve army, the direction-of-innovation argument and the compensation critique while declining the labour theory of value that, for Marx, grounded all of them. The defence — that Engels stated the reserve army in 1845 with no value theory attached, and that modern search-and-matching is its descendant — is good, and it is the strongest single move in §4. But a Marxist referee will say the paper has taken the sociology and left the economics, and they will not be entirely wrong. 2. **The Fragment on Machines is an unpublished notebook Marx abandoned.** The paper says so, cites Heinrich, and uses only the weak reading. Still: the most rhetorically effective passage in §4 comes from a text the author concedes its own writer walked away from. Expect that sentence to be quoted back. 3. **Two pillars can look like hedging.** If Polanyi fails the test in §11 and Marx does better, why is the paper still called *The Third Disembedding*? There is a real answer — the disembedding index is the measurement instrument regardless of whose story it serves — but the title now oversells the Polanyian half. **The limit-case section is the strongest new material and has one real weakness.** 4. **Proposition 2 is a property of a functional form.** The three cases fall out of a CES nesting; no elasticity has been estimated, and the paper says so. The result is genuinely useful — it identifies precisely which unmeasured parameter each rival conclusion depends on, and shows the leading model assumes its own answer — but it is a clarification, not a finding, and §8.4's table should be read as the section's actual output. 5. **The critique of Trammell and Korinek's Cobb–Douglas nesting is fair and slightly unfair at once.** Fair: the resource share is pinned by the functional form. Unfair: they say explicitly that the fixed-factor case is a robustness exercise, not their main model, and their central argument does not depend on it. State that. 6. **"Land takes everything" is a very old result.** Ricardo has it; George has it; the paper reaches it with more machinery. The novelty is the *route* — deriving it from the reproducibility of capital in an automation model — not the destination. Say so before a referee does. --- ## 7. What the paper gets right, and should not trade away Listing these because the revision pressure will be to fix the weaknesses by softening the strengths, and that would be the wrong trade. - **It reports its own adverse result in the abstract.** Very few papers do this. It is the single most credible thing about the draft. - **The claims-and-citations box.** Stating up front what the paper may not be cited for is the right response to writing something quotable about a contested topic. - **The Speenhamland concession.** Most Polanyi-invoking work does not know the historiography exists. - **Crafts is engaged rather than avoided.** The leading contrary case is stated at full strength, in the text, with its numbers. - **Fraser's warning is applied against the paper's own implied politics.** The observation that a professional countermovement would most likely take the form of licensure — regulatory capture in Polanyian costume — is the paper's sharpest paragraph and cuts against its own sympathies. - **Forecasters are cited as positions and never as evidence**, with financial interests disclosed. This alone puts it ahead of most of the AI-economics commentary literature. - **The measured-versus-projected discipline** in §13 and Figure 11. - **Draft 2.0 ran a second, independent verification pass and it found that the first pass had corrected a figure in the wrong direction.** That is the right argument for doing verification twice with different framings, and the paper should keep doing it. --- ## 8. Priority list for draft 1.1 **P0 — do these before showing anyone who matters.** 1. **Run Study A.** Everything in §10 onward is a promissory note until the experience–wage gradient is estimated. It is a day's work on public data. Until it exists, retitle §10 from "The recovery" to "A conjecture" and move the strong claims into the conditional. 2. **Demote Proposition 1 from theorem to accounting identity.** State plainly that it is the derivative of a conditional mean and that its value is in telling you what to measure. Losing the false grandeur costs nothing and removes an easy hit. 3. **Reduce the τ series to T3, or add a standing caveat to every use of it.** Two of three response dates are triggered by something other than the technology, and the paper says so once while the figure says otherwise nine pages later. **P1 — before circulation.** 4. Study B: the six-digit reconstruction of the protection gradient. Or state the current estimate as "no detectable relationship" throughout, which is what n = 11 with those intervals supports. 5. Answer the Gemici (2015) "what does Polanyi buy you" question in a paragraph, or accept that the framing is presentational. 6. Drop J51. 7. Add within-group heterogeneity to the discussion of why occupation groups are a poor proxy for tasks. **P2 — nice to have.** 8. An appendix deriving the wage expression rather than citing it, so §6.1 earns its place. 9. A robustness table: the two correlations under alternative exposure measures (Felten–Raj–Seamans, Eloundou, Anthropic), alternative protection measures (member vs represented, licence vs licence-or-certification), and weighted vs unweighted. 10. Sensitivity of ρ* to the three free parameters of the response function, so the reader can see how little it means. --- ## 9. The bench score | Dimension | Score /10 | Note | |---|---|---| | Honesty about own weaknesses | **9** | the best thing here; adverse result in the abstract | | Historiographical care | **8** | Speenhamland and Crafts handled properly | | Citation hygiene | **8** | post-verification; two hazards flagged in the text itself | | Novelty of the framing | **7** | expertise-with-a-bearer is a real answer to Kenney et al. | | Falsifiability apparatus | **7** | the death table and Prediction 1 are genuine | | Empirical support for the central claim | **3** | n = 11, null result, vendor telemetry | | Forcing power of the formal model | **3** | Prop. 1 is an identity; the production block is inert | | Calibration of τ and λ | **3** | three anecdotes, two with the wrong trigger | | Marx sections: accuracy and edition discipline | **8** | quotations located, translations distinguished, Heinrich conceded | | Limit-case derivation | **7** | correct, clarifying, and honest that it identifies parameters rather than estimating them | | **Overall as a framework paper** | **7** | up from 6.5; the second pillar closed the "no engine" hole | | **Overall as an empirical contribution** | **3** | unchanged. Study A still has not been run | **Verdict slogan:** *The defence wins the character of the paper. The attack wins the evidence. Run Study A, demote Proposition 1, and the kill shot stops landing.* **Draft 2.0 addendum.** The theory is now in better shape than the data by a wider margin than in 1.0. Adding Marx fixed a real hole — Polanyi had no engine — and the limit case answers a question the framework kept raising. Neither addition moves the empirical score, and the gap between a 7 on framework and a 3 on evidence is itself the criticism. One day of CPS work would do more for this paper than another ten pages of theory. --- --- # ATTACK NOTES — *Nobody to Blame*, draft 4.0 Filed 2026-08-01. Draft 4 abandons draft 3's thesis. The referee below is the same hostile competent labour economist, now with a political scientist sitting next to them. --- ## The kill shot > *The author has built a theory of political attribution on a single survey experiment > with hypothetical vignettes, and then used it to interpret a decade of actual > elections. Stated preferences over a fictional plant closure are not enacted trade > policy. Every observational result the paper cites — the robot-shock voting > literature, the China-shock realignment literature — was designed to test magnitude, > not form, and is therefore consistent with the theory only because it is silent on > it. The "reconciliation" reconciles nothing: it names a distinction the data cannot > see and declares the disagreement dissolved.* **Status: this one lands.** The paper's defence is that the distinction is stated as a falsifiable prediction (P3, P4) with resolution dates in 2026 and 2028, and that the Di Tella–Rodrik compensation-margin sign flip is not something a magnitude story predicts. That is a real defence and it is not yet a sufficient one. Section 4's caveat box says so; the disclosure says so again. Until P3 or P4 resolves, the central claim is a well-motivated conjecture. ## Second: the entry-level evidence is not an AI result and the paper knows it Emanuel, Harrington and Pallais attribute **64%** of the rise in young-graduate unemployment (2017–19 → 2022–24) to remote work, *through the same lost-mentorship mechanism this paper describes*. That paper concedes the mechanism and takes the attribution. Add that the recent-graduate/overall unemployment crossover is **February 2019** — pre-pandemic, pre-AI — and the honest reading is: the entry rung is under strain, the mechanism is mentorship, and AI's share of it is unidentified. The paper says all of this in a box in §7. Saying it does not make the empirical foundation stronger; it makes the paper harder to accuse of hiding it. Those are different things and the referee should press on the first. ## Third: the pipeline has no case behind it No documented instance of a training pipeline collapsing **while demand held up**, with clean numbers and a decade-lagged consequence. UK apprenticeships halved 1979–95 but the industries died first — causation runs the wrong way. So the argument is: a theorem about equilibrium multiplicity (solid), a measured mentorship cost of 0.76 programs per month (solid, one firm, one occupation), and a claimed regime change in whether training is free (**unmeasured, and the paper's own contribution**). Indicator 6 — training spend per junior hire — is the thing that would settle it and no time series exists. That is a genuine hole and not a rhetorical one. ## Fourth: "the bin is stable, its contents are not" is doing a lot of work The junior-posting evidence is flatly contested. Indeed says entry postings −7.5% y/y against senior +15%; two Lightcast-based studies find junior and senior demand moving in parallel. The paper resolves this by asserting that titles are the wrong unit and task text is the right one — supported by PwC (entry roles demanding senior skills +35%, non-seniorised entry roles −10%) and Indeed (30% of entry-level applications from people with 10+ years' experience). That is suggestive. It is not the study, and the study has not been run. Indicator 4 states the test; nobody has executed it. ## Fifth: the feedback loop is stated in general and contradicted in particular Protection raises the price of labour and induces automation — Bracero (Clemens–Lewis– Postel) and ageing (Acemoglu–Restrepo, ~40% of cross-country robot variation) are clean. But the *enacted* 2024 Section 301 schedule puts 25–50% duties on semiconductors, batteries and solar: tariffs on **capital goods**, which raise the price of automation. Flaaen and Pierce put the 2018 round's net manufacturing employment effect at **−2.7%**. So the mechanism is well evidenced and its application to this tariff schedule is ambiguous in sign. The paper concedes this in §6 rather than burying it, which is correct and also removes the section's punchline. ## Sixth: five sections of philosophy in a paper about JOLTS Reduced sharply from draft 3 — the Locke material survives only as §10 in the full edition and two sentences in §12.4. The referee will still ask what it is doing. The answer the paper gives: the limit case reverses Locke's explicit empirical premise, and the contractarian exclusion problem is the philosophical shape of the political worry in §12.4. Thin, but not decorative. --- ## What would make the paper unattackable on its own terms 1. **Run indicator 5.** Within-occupation return to experience by AI exposure, CPS or ACS, a day's work. If the gradient compresses differentially in exposed occupations and concentrates in the first decade of tenure, the expertise-rent mechanism has an identified test rather than a field experiment in one occupation. 2. **Re-run the posting analyses on task text.** Indicator 4. The data exists; the analysis does not. 3. **Find one historical pipeline collapse with demand intact.** If it does not exist, say the mechanism is novel rather than analogous — which is a stronger claim and a more exposed one. 4. **Wait for November 2026 and November 2028.** P4 resolves on real votes. ## Errors corrected in this draft - Draft 3's central claim (λ ≈ 0 since 1980) — retracted in §3 and in the disclosure. - "88% of manufacturing job loss was automation" — this is a residual in Hicks–Devaraj's accounting decomposition, not an estimate. Houseman's critique on the measurement is decisive: computers are 10–13% of manufacturing value added, and excluding them, manufacturing real value added was *lower* in 2011 than in 2000. - "Technology escapes political blame" — contradicted by Anelli–Colantone–Stanig and by the fourteen-country horse race. Reframed as form-not-magnitude. - "The entry-level adjustment is invisible to official statistics" — wrong. It is visible in BLS unemployed new entrants. What is invisible is underemployment. - METR follow-up is **−18%**, not +18%. A widely shared 2026 summary has the sign error. --- ## Draft 4.1 update (1 Aug 2026) - Kill-shot #3 ("pipeline has no case") is now half-answered: PATCO 1981–92 gives a measured 11–14-year repair time under maximal effort (reverse experiment). The referee's residual line: "a repair-time bound is not evidence collapse happens." Correct; §8.1 says so itself. - New exhibit for the hiring-margin claim: NY WARN AI-disclosure read zero across 160+ notices in year one. Referee will say the checkbox is voluntary-ish and undefined — conceded in text. But note the jiu-jitsu: the paper *predicts* zero on separation-triggered instruments, so this cuts for us either way it's read. - §7 cuts-against box now carries EPI's noncollege-parallel objection and the NY Fed's pre-ChatGPT trend point by name. The box is four-deep now; the section survives because it claims mentorship strain, not AI causation.