Synthesis — Korinek, Jones, Sacher, Cotter & McCrory (2026), "Economic Scenarios for Transformative AI"

Replication: OPEN · 2 not established / 15 narrower / 12 hold

Status

COMPLETE. No claim-ledger row is INTERIM, nothing waits on further work, and no computation tied to the job is running (checked 29 Oct 2026).

Replication is OPEN, not TRUSTWORTHY: every script a load-bearing number rests on re-ran with identical output and matches the reports and, where the paper prints the number, the print, but some checks are blocked outside the job. Table 4's per-respondent runs need survey microdata that were not deposited, and the sources listed next are unverified here.

One peripheral row, the benchmark sentences on pp. 3 and 33, rests partly on works the review could not read in the form the paper cites. They are unverified-here after every acquisition tier, operator access included (00_INTAKE.md, acquisition notes): a limit on the review, not a finding about the paper.

• Acemoglu (2025) was read as NBER w32487 (May 2024), which the paper's references list beside the Economic Policy 40 (121) article, and Karabarbounis and Neiman (2014) as NBER w19136 (revised October 2013), not the QJE 129 (1) article. The row's figures for these two works are the working papers': Acemoglu's TFP bounds (0.66 and 0.53 percent) and GDP ranges (0.93–1.16 and 1.4–1.56 percent, over ten years), and Karabarbounis and Neiman's 1975–2012 sample, 5-point global decline and US series (their Figure 2).
• Filippucci, Gal and Schief's AEA Papers and Proceedings article (cited as 2025, forthcoming; indexed as vol. 116, 2026): the publisher's full text is restricted and only the abstract record was read. The row's figure for it is the abstract's "up to 0.9 percentage points" of annual TFP growth.
• The Penn Wharton Budget Model brief read is "The Projected Impact of Generative AI on Future Productivity Growth" (8 Sep 2025). The paper cites "Projecting the economic effects of artificial intelligence"; no item under that title was found, and the operator confirmed the brief as the cited item. The title mismatch is noted wherever the brief is used.

A figure changed in a published version would change that row's numbers. The body text of CEPR DP21939 is also unverified here (see Version of record); no row rests on it.

Every challenge ruling relied on below was re-run from its script, gave identical output, and had its page re-opened: for the three original challenge groups in analysis/challenge-rerun-check.md; for the challenge of the two Introduction rows on cognitive wages and on capital's return and share, in work/verdict/rerun/ (its sign scan, variants, relative wages and exits-routing copy, each identical, the files they rewrite unchanged; Table 3 on p. 31, the p. 4 paragraph and Figure 3's cognitive-wage panel on p. 33 opened as images). None was set aside. The audit-built calculations cited here were re-run with identical output: the benchmark-literature comparisons (work/verdict/rerun/benchmark_lit.log), the compensation identity applied to the review's saved closures and grid (work/verdict/rerun/fix_comp_nowrite.log, every recomputed section) and the Figure 3 wage scan (work/verdict/rerun/aud_fig3_wages.log).

Version of record. The Anthropic Institute Working Paper No. 2026-02, September 2026 (title page, p. 1). It is byte-identical to the copy on Charles Jones's site and has not been revised since its release on 9 Sep 2026 (00_INTAKE.md; analysis/paper/drift-table.md). The earlier draft acknowledged on p. 1 is unobservable, and the body text of CEPR DP21939 is unverified here.

Deposit: partial.

• What it holds. The scenario explorer's compiled model code (04_replication_korinek_jones_2026_explorer_kernel/0t5oquk5tpzvc.js, the authors' public implementation, fetched 27 Oct 2026) and a snapshot of the explorer page. The operator assembled it; it is not an author replication package.
• What it lacks. The calibration code (how the Table 1 and Table A.2 targets were built from CPS, IPUMS, JOLTS and AI-usage data). The survey microdata, weights and coding code. The code that printed the tables and figures.
• Unverifiable as published, therefore:
  • Table 4's per-respondent quantiles, and Table 2's medians beyond their printed cells.
  • Whether Table 4 used the explorer's survey mappings.
  • How the mid-2026 observed-exposure anchor was built (0.14; a third-party recompute gives 0.1212).
  • Which code printed Tables 3, 5 and 6. The explorer reproduces them, which shows they are its outputs, not that it printed them.

Conflict of interest. This review was run with Claude, made by Anthropic. The paper is an Anthropic Institute working paper, and its authors say they "used Claude as a research and writing assistant" (p. 1, first footnote). The review is AI-generated input for The Unjournal's commissioned human evaluations. It is not an Unjournal evaluation and has not been reviewed by a person. The two earlier AI reviews of the paper in lit/ were used only as leads; nothing here rests on them.

What the paper establishes and what it does not

What it establishes. The paper's model is implemented as printed, and its tables are the model's outputs. The authors' public scenario-explorer code:

• reproduces every printed cell of Tables 3, 5 and 6 to the printed rounding;
• has not changed since September;
• solves the stated model at all 7,776 points swept.

Proposition 1 re-derives from the primitives (review/code-and-data-REPORT.md, findings C1, C14, C17; work/challenge/scenarios/s3_anchor.log:22; work/challenge/compensation-theory/s2_kernel_battery.log:157).

The paper says its scenarios are conditional, not predictions (p. 3). Read that way, the modest and extreme numbers hold:

• Modest change. In 2030, GDP growth is 0.40 points higher and all-worker unemployment 0.06 points higher (3.90 against 3.84). Both survive every alternative common parameter the paper uses (work/challenge/scenarios/s3_anchor.log:17; work/challenge/scenarios/s4_variants.log:34–45).
• Extreme change. GDP is 32.4 percent above its no-AI path, growth is 15.4 percent, and the labor share falls to 45.2 percent (Table 3, p. 31). "Transformative" survives every combination of the paper's own parameter values: GDP is 15.7 to 52.9 percent above the no-AI path (work/challenge/scenarios/s5_grid.log:80). The size moves mainly with the capital-supply elasticity, as Table 5 discloses.

The directions of the capital-supply and wage-rigidity results hold cell for cell (Tables 5 and 6). In all three scenarios the return to capital, the capital stock and the capital share are above their no-AI values in every month from 2024 to 2030, and the labor share falls, as pp. 3–4 say (work/challenge/new-rows/s3_scan.log:36–40, 90–94, 144–148).

With the task-grid capability measure, the survey implies a 2030 economy nearest the substantial scenario in GDP and cognitive employment. That holds for both computations the paper uses: Table 4 gives 8.6 and −4.2, and the model at Table 2's item medians gives 9.9 and −4.5. It also holds under every reading of the re-employment answer tried and under random item non-response. Under no variant does the survey land near the extreme scenario.

The extreme scenario's gains could compensate cognitive workers. Holding them at their no-AI income leaves the rest of the economy ahead on every basis consistent with the model's capital supply, by 6.3 to 32.4 percent of no-AI GDP at the scenario's own parameters (work/challenge/compensation-theory/s2_kernel_battery.log:164–172, 207–227). With the extra capital foreign-owned at the model's rental, the rest stays ahead in every one of the 1,296 extreme-capability combinations of the paper's own parameter values (+1.4 to +15.5; work/verdict/grid_paper_values.log).

Under the modest or substantial scenario the reallocation is, as p. 39 says, of a size the U.S. labor market has absorbed. The substantial scenario lowers the cognitive share of employment 2.0 points in 3.5 years; documented broad-group shifts, at their decade-average paces, would move a share 1.2 to 2.3 points in that time. Its 2030 unemployment, 4.6 percent, is well inside past peaks (work/interim/reallocation_vs_history.log:3, 9–16; p. 35). In points a year it is at the fast edge: only agriculture's 1900–30 decline was faster, and no 3.5-year historical window was obtained.

What it does not establish.

How the survey enters the model (design-level finding D1). This narrows the most headline claims.

• The abstract and the Introduction say the survey lands close to the substantial scenario. That holds with only one of the two capability measures the paper prints. The other is employed cognitive workers' estimate of the share of their own hours AI can do (p. 47). With it, every outcome lands nearest the modest scenario (GDP 3.7 percent), and the points where the result flips lie between the two measures.
• The p. 30 claim that respondents' eight months is "slower than the substantial scenario assumes" is not established. The mapping makes the questionnaire's "three months" stand for a 27.7-month wait for a job in the other group, in the model's own units. Under every model-consistent reading, eight months is faster than the substantial scenario implies.
• Whether the survey's unemployment looks like the substantial or the modest scenario turns on what respondents meant: total time to a new job, or time to a job in a new occupation. The data cannot say which.

"Nearly one in five cognitive workers is unemployed" (D2). This is the extreme bundle's correct output. But its size rests on a premise nothing in the paper anchors: that the people who replace cognitive workers leaving the labor force search as cognitive-origin job seekers.

• Across routings that keep the calibrated steady state, the figure is 13.9 to 17.6 percent. It is in the teens always, and "nearly one in five" only if few of those replacements enter the growing occupations.
• The same premise carries the substantial scenario's "more than a 50 percent" rise in cognitive unemployment. Across the routings that rate is 2.0 to 4.5 percent in 2030. As specified, the rise from the scenario's own mid-2026 rate is 48 percent.
• It also carries the size of the extreme wage split, and the substantial clause of the Introduction's sentence that cognitive wages grow more slowly than without AI.

"The aggregate gain is nearly three times what cognitive workers lose" (p. 39; D3). This is a gross-output multiple.

• The extreme scenario installs extra capital worth 1.96 times no-AI GDP. Once that capital is charged for, the gain is 0.54 to 1.95 times the loss.
• On the national-income basis that the paper's own world-market financing rationale implies (p. 26), the aggregate gain is about half the loss. It is still positive, which is all compensation needs: paying the loss leaves the rest of the economy the whole net gain.
• The resources therefore exist as potential compensation. They do not exist as consumable 2030 output if the build-up were financed at home. And p. 39 drops p. 35's "in the textbook sense".
• The same identity removes p. 37's reading of footnote 14's transfer: "compensating workers would consume most of the gains" holds only as a comparison of sizes (the transfer is 84 to 87 percent of the GDP gain). After the transfer the gainers keep the whole gain.

Timing and the innovation channel.

• "Almost all of the divergence comes after 2027" is about four-fifths at the date the paper names.
• The scenarios share only two of the five AI readings they are said to share.
• The small 2030 innovation effect is labelled a "surprising finding", but it follows from the research-funding rule, which the paper itself derives (pp. 16–17).

Beyond the headlines.

• The Introduction's "in each scenario ... cognitive wages grow more slowly than they would without AI" (p. 4) fails for the modest scenario as specified: its cognitive wage is above the no-AI path in every month from 2024 and 0.4 percent above it in 2030, as the paper's own Table 3, Figure 3 and p. 32 print. The clause holds for the substantial and extreme scenarios as specified. Under the exits routing it holds at 2030 only for the extreme.
• The claim that coming data "will tell us which scenario we are in" has no test.
• The benchmark sentences (pp. 3, 33) hold as levels, or as a pace from 2024, not on the cited works' horizons, and not for every work cited. The modest scenario's 1.6 percent of GDP by 2030 is slightly above Acemoglu's upper ten-year range but 1.4 to 1.7 times his baseline range, and it lies below every OECD figure it is said to be "of the same order as" (they are 1.3 to 7.6 times it). Briggs and Kodnani's pace runs for ten years from widespread adoption, which they do not date. Of the other works said to give "comparable magnitudes" to the substantial scenario, Aghion and Bunel do, while the Penn Wharton brief puts 2030 GDP 0.6 percent above a no-AI path, one-fourteenth of the substantial scenario's 8.3 and below the modest one's. Karabarbounis and Neiman's data cover about three decades after 1980, not four, and the corporate sector only.

Design-level findings

D1. How the survey's answers enter the model: the capability measure, the reading of re-employment time, and the sample

bears on: the abstract's survey sentence (CL06), the Introduction's survey sentence (CL07), the survey results paragraph (CL14), the eight-month ordering (CL15)

How the comparison is built. The paper turns five survey items into five model inputs and fixes four other inputs at the substantial scenario's values (Table 4 note, p. 36). It computes the comparison two ways:

• The model at Table 2's item-by-item medians. This is the Introduction's sentence. It is never printed. It gives GDP 9.88, cognitive employment −4.49, cognitive unemployment 5.16 and overall unemployment 4.89. The explorer's quiz gives the same.
• Table 4's medians of per-respondent runs. These cover the 3,259 respondents who answered all five items, with survey weights. They give 8.6, −4.2, 4.6 and 4.6 (work/challenge/survey/s4_core.log:8–9; p. 36).

Three choices decide where the survey lands.

Capability. The task-grid answer (m = 0.44) is one of two measures the paper prints.

• Employed cognitive-occupation respondents put the share of their own hours AI can do by 2030 at a median of 27 percent, against 71 from the grid (p. 47).
• Scaled the same way, that is m = 0.160 (0.168 without the 0.95 in-person factor). At that value every outcome is nearest the modest scenario: GDP 3.70, cognitive employment −1.15, cognitive unemployment 3.31, overall 3.97.
• The result flips between m = 0.2175 and 0.2575, which lies between the two printed measures (work/challenge/survey/s4_core.log:10–16; code-and-data C5; assumption ledger WA03).

Re-employment. The rule μ = 0.17 × 3/months maps the questionnaire's "three months" to the normal-times search discount. At that discount, the model's cognitive-origin worker waits 27.7 months for a job in the other group, and each Table B.1 code stands for 8 to 15 times its months (s4_core.log:22; s7_extra.log:1–9). Two readings of the median answer give different results:

• Time to a cross-group job. Survey unemployment lands at the modest scenario's level (cognitive 2.9–3.4).
• Total re-employment time. This is the object question 5 names ("how long ... to find a new job", p. 46). Survey unemployment stays nearest substantial (cognitive 4.2–4.4, overall 4.5) (s4_core.log:47–50; s7_extra.log:10–15).

Under every model-consistent reading, eight months is faster than the substantial scenario implies (s4_core.log:26–38; s7_extra.log:16):

• under cross-group readings, faster than all three scenarios;
• under the total reading, between the modest scenario (3.2–3.5 survey months) and the substantial one (9.1–9.8).

GDP and cognitive employment stay nearest substantial under every reading (GDP 10.3–10.9).

Sample. Table 4 uses 29.7 percent of the 10,980 respondents, and the paper states no assumption about item non-response.

• At the corners the print allows, the median economy runs from GDP 1.8 to 31.7. This is a worst-case bound, not an estimate (s4_core.log:74–75).
• Under random item non-response, synthetic respondents matched to every Table 2 quartile give medians of GDP 6.5–10.2. All 29 runs, across 20 dependence structures, are nearest the substantial scenario on all four outcomes (s5_synth.log:6–20; s6_synth_seeds.log:6–19).
• No synthetic run reproduces Table 4's GDP quartiles [3, 19], so the print does not reconcile its two survey tables. That gap depends on how the synthetic sample is built and is not evidence of selection. Table 4's note says it uses survey weights; Table 2's does not say.

Choices that move the level but not the nearest scenario.

• Fixing reinstatement and posting speed: every combination of their scenario values stays nearest substantial (GDP 8.0–11.7; s7_extra.log:17–25).
• The productivity answer's code edge, which the paper notes (p. 49): GDP 6.2–12.2 (s4_core.log:55–72).
• Reading the productivity question's present tense as today's gain: GDP 11.0 (s4_core.log:73).

Choices that are fragile or unstated.

• Adoption. The share of capable tasks people "will regularly use it for" is coded directly as the share of instances done with AI. This is fragile: the median run moves nearer modest once that share falls below 0.1975, an instance share of under about half among regularly used tasks (s4_core.log:17–20; assumption ledger WA04).
• Coding. The explorer's quiz maps "less than 1 month" to μ = 0.51, against Table B.1's printed 1 (s2_helpers.log:2, 34). It also runs answers below today's AI levels at the mid-2026 anchors (code-and-data C26). Whether Table 4 used either rule cannot be checked.

What this removes.

• The eight-month ordering on p. 30 is not established.
• The survey's closeness to the substantial scenario is narrower in the abstract, the Introduction and the results paragraph. It holds with the task-grid capability measure, not with the own-hours measure the paper also prints. Its unemployment component holds only if respondents meant total re-employment time.
• For US adults as a population (Table 4's complete-case sample) the print cannot say, though random non-response does not move the result.

What survives. With the task-grid measure, the survey lands nearest the substantial scenario on GDP and cognitive employment: under every reading, every fixed-input setting and random non-response. Under no variant does it land near the extreme.

D2. Where the replacements of cognitive workers who leave the labor force search for work

bears on: "nearly one in five cognitive workers is unemployed" (CL05), the mechanics sentence (CL10), the substantial scenario's rise in cognitive unemployment (CL13), the extreme wage split (CL17), "cognitive wages are slightly lower in 2030 than in mid-2026" (CL26), the Introduction's cognitive-wage sentences (CL28)

How the model routes exits.

• All separations pass through the unemployment pool (p. 21).
• Exits from the labor force are 0.05 of the 0.11 annual quit rate (Table A.2, p. 44).
• Every quit from a cognitive job enters the pool as a cognitive-origin searcher, who crosses to the other group at the discount μ (paper report P9).
• In the modest and substantial scenarios no worker is laid off in any month (work/challenge/scenarios/s3_anchor.log:24, 26). Cognitive employment falls only through quits that are not replaced. So the substantial scenario's whole rise in cognitive unemployment is the search of quitters, including the people who replace retirees.

The rebuilt routings. The challenge rebuilt the routing on the authors' kernel as a family of rules that keep the calibrated steady state. With the switches off it reproduces the kernel; the labor-force identity holds to within 9 × 10⁻¹⁶; nothing is routed when AI is off (work/challenge/scenarios/s6_routing.log:1–17, 50–53; s7_routing_spillovers.log:2–7). A second copy, built for the cognitive-wage sentences, reproduces the same unemployment and extreme-wage figures and adds the substantial cognitive wage (work/challenge/new-rows/s7_routing.log:3–15, 29–50).

Wages and other outcomes.

• The extreme wage split is −11.54 and +33.61 as specified (cognitive and other wage against the no-AI path). It moves to −2.71 and +22.23 with all replacements redirected, and to −8.50 and +29.66 with a quarter.
• Across the rules that send replacements to the growing group, the cognitive wage is 1.9 to 8.5 percent below its no-AI path and the other wage 21.3 to 29.7 above. Under proportional entry they are −11.18 and +33.05. The extreme cognitive wage therefore stays below its no-AI path in 2030 under every routing, though above it in 4 to 32 of the 61 months from 2025 under the redirecting rules (work/challenge/new-rows/s7_routing.log:44–50).
• The substantial cognitive wage ends 1.03 to 1.45 percent above its no-AI path once a quarter or more of the replacements are redirected, and is above it in 37 to 47 of 61 months (s7_routing.log:33–38).
• The modest cognitive wage is above its no-AI path in 2030 under every routing that passes the AI-off check (+0.42 to +0.60; s7_routing.log:18–28).
• Extreme GDP rises (36.41 with all replacements redirected), and the labor share stays at 44.96–45.21.

The premise is FRAGILE, not WHIMSICAL (review/challenge-REPORT-scenarios.md, element 19).

• Its effect is a threshold: inert under proportional entry, material when replacements are sent to the growing group.
• Its warrant is unanchored. The paper's entrant rule (p. 20) points to proportional entry. The human-capital rationale for the search discount (pp. 19–20) points the other way for first-time entrants, as does p. 21's remark that past adjustment ran through entry and exit.
• Nothing in the packet measures the share.
• Sending all replacements to the growing group (14.04 and 3.06) is the strong end of the family, not a supportable central figure.
• With AI switched off, almost nothing is routed, yet every rule leaves 2030 cognitive unemployment at 2.88–2.90 against the no-AI path's 2.85 (analysis/challenge-rerun-check.md, observation on the AI-off check). The family's levels carry that offset of 0.03–0.05 points; the cognitive wage's AI-off offset is at most 0.016 points under the rules used here (work/challenge/new-rows/s7_routing.log:9–15).

What this removes. The following rows narrow, on the extreme and substantial scenarios:

• "Nearly one in five" as the extreme figure, once a quarter or more of the replacements enter the growing group.
• "More than a 50 percent increase" in substantial cognitive unemployment.
• The stated sizes of the extreme wage split.
• The sign of the cognitive-wage change from mid-2026 to 2030.
• That substantial cognitive wages "grow more slowly than they would without AI" (p. 4).

What survives.

• Extreme cognitive unemployment in the teens, "far above any postwar rate in the United States" (p. 35), under every routing.
• A steep fall in the relative cognitive wage, and an extreme cognitive wage below its no-AI path in 2030, under every routing.
• Extreme GDP and the labor share, unchanged or higher.
• The mechanics sentence. It says unemployment "depends both on" the shock's size and switching ease, not only on them.

D3. Charging for the capital the extreme scenario installs, and what compensation leaves

bears on: "the resources to compensate them exist" (CL08), the textbook-compensation passage (CL16), footnote 14's "compensating workers would consume most of the gains" (CL18)

The capital and the GDP-basis numbers.

• In 2030 the extreme scenario's capital stock is 56.3 percent above its no-AI path (Table 3, p. 31). At the no-AI capital-output ratio of 3.48, that is extra capital of 1.96 times no-AI GDP (code-and-data C3).
• The model has no consumption-investment split (footnote 5, p. 16) and no Euler equation (footnote 13, p. 35). It justifies ε = 3 by compute "financed in a world market" (p. 26).
• On the GDP basis, the 2030 gain is 32.41 percent of no-AI GDP against a cognitive wage-bill loss of 11.61: 2.79 times. The transfer is 8.8 percent of AI-economy GDP, the "about 9 percent" p. 35 prints (s2_kernel_battery.log:160).

What compensation leaves. Before the transfer, the rest of the economy's change is the gain plus the cognitive loss (44.02). Paying the loss leaves exactly the gain (analysis/challenge-rerun-check.md, part (i); s2_kernel_battery.log:171). So holding cognitive workers at their no-AI income leaves the rest ahead by the whole net gain, on whichever basis the gain is measured and whatever the transfer's size. The multiple of gain to loss measures how large the transfer is relative to the gain, not whether compensation is possible: that needs only a positive net gain.

At the scenario's own parameters (s2_kernel_battery.log:164–172):

Across the paper's parameter values.

• In all 32 cells of the Table 5 × Table 6 grid, both scenarios, the rest is ahead after the transfer on every basis that charges for the capital (foreign-owned minimum +2.0), and behind in every cell on the domestic 2030 flow (s2_kernel_battery.log:207–227).
• Over the 1,296 extreme-capability combinations of the paper's own values of the search discount, posting speed, automation share, reinstatement, capital-supply elasticity and wage rigidity, the rest is ahead in every cell with the extra capital foreign-owned (+1.36 to +15.48) and behind in every cell on the domestic 2030 flow (−137.6 to −7.9). "More than 20 percent ahead" holds in 89 percent of the cells on the GDP basis, 52 percent on net product, 29 percent on sustained replacement and none with the capital foreign-owned (work/verdict/grid_paper_values.log). These are the paper's-values cells of the review's joint grid, recomputed with the identity from its saved outputs (work/audit/fix_comp.log section C, whose 2,592 cells add a steeper logistic the review introduced). Their GDP-basis range, 15.66–52.87, is the one the scenarios challenge's own grid gives (work/challenge/scenarios/s5_grid.log:80).

The one closure that fails. Only financing the build-up out of 2030 domestic output fails. The rest stays ahead if no more than about 7 percent of the build-up is financed that way. That is the closure p. 26 sets aside: there, ε = 3 is "higher than five years of saving at historical rates would imply" (s2_kernel_battery.log:162, 168–170).

At ε = 1 the rest is +21.3 on the GDP basis and +5.6 on the supply-price basis (s2_kernel_battery.log:234).

The wage-bill share. The cognitive occupations' share of the May 2025 OEWS wage bill, on footnote 10's grouping, is 0.6933 (pay ratio 1.77). The model uses the employment share, 0.624, for both (p. 25; Table A.2, p. 44) (work/challenge/compensation-theory/s3_oews.log:47–50).

• The rest's position after the transfer equals the gain whatever the share, so no wage-bill share makes the cushion fail.
• The multiple stays at or above 2.5 on every construction built: 3.51–3.68 when the measured share is run through the model, a variant that moves the employment and wage-bill shares together, and 2.51 when the loss is scaled with the share at a fixed percentage loss (s5_share_variant.log:4).
• The transfer's size is not settled. The cognitive loss to be replaced is 9.2 to 11.0 percent of no-AI GDP in the model variants and 12.9 under fixed-percentage scaling, against 11.6 at the paper's share (s5_share_variant.log:2, 4). A version with heterogeneous pay, which would settle it, was not built.
• The GDP-basis multiple also moves with the across-task elasticity σ, which the paper does not vary: 1.58 at σ = 0.25 and 5.38 at 0.95 (code-and-data C11).

The premise labels (analysis/code-and-data/assumption-ledger.csv, rows on the transferable GDP gap and on the wage-bill share), with the rest's position after the transfer equal to the gain:

• The GDP gap as a pool of transferable resources is WHIMSICAL, and its effect is MATERIAL, not a sign flip. Under the ledger's central relaxation (the extra capital priced along the model's supply schedule) the rest stays ahead (+8.0). What moves outside the paper's wording is the multiple (0.69 against "nearly three times") and the cushion (8.0 against "more than 20 percent"). The sign flips only at the pole, domestic financing of the whole 2030 build-up, which p. 26's financing rationale sets aside. The warrant is a stated convenience (footnote 5). The ledger's minimal repair, as its note states it, is what the compensation rows can carry.
• The wage-bill share set to the employment share is not FRAGILE: no threshold exists. The cut-offs at shares of 0.667 and 0.696 arise only when the loss is counted twice or scaled at a fixed percentage loss (review/challenge-REPORT-compensation-theory.md, elements 10, 15 and 38). The OEWS measurement contradicts the premise, but no construction built breaks a conclusion it carries: it is inert for the cushion and the multiple, and unresolved for the transfer's size (ledger ruling UNRESOLVED).

Three printing points.

• p. 39 drops p. 35's "in the textbook sense".
• p. 35 does not print the base of "more than 20 percent ahead". The figure 32.4 − 11.6 = 20.8 is the one that rounds to the print. The rest's actual position (32.4 percent of no-AI GDP; 51.8 percent of its own income) is also "more than 20 percent".
• p. 37 reads footnote 14's transfer, "84 percent of GDP gains", as "compensating workers would consume most of the gains". At the footnote's parameters the gain is 7.21 percent of no-AI GDP and the transfer 84.2 percent of it on mid-2026 employment, 87.0 on no-AI-path employment (s2_kernel_battery.log:241–243). After the transfer the gainers keep the whole 7.21. The transfer is 46 percent of their gross gain of 13.5 (87.0 percent of 7.21 is 6.27; 6.27 of 7.21 + 6.27; analysis/challenge-rerun-check.md, qualification 4).

What this removes.

• "Nearly three times" (p. 39) and "almost three times" (p. 35) as measures of the resources available once the capital is charged (0.54–1.95).
• "More than 20 percent ahead" on the replacement, supply-price and foreign-owned bases, and on the foreign-owned basis in every cell of the grid.
• The resources as consumable 2030 output under domestic financing.
• "Compensating workers would consume most of the gains" (p. 37) as a statement of what compensation leaves.

The resources sentence, the textbook-compensation passage and the so-so example are therefore narrower.

What survives.

• Potential compensation, on every closure consistent with the model's capital-supply premise, at the scenario's own parameters and across the paper's parameter values.
• The GDP-basis multiple of 2.79 and the 8.8 percent transfer.
• "More than 20 percent ahead" on the GDP and net-product bases.
• For footnote 14, the transfer's size against the net gain: 84.2 percent on mid-2026 employment, 87.0 on no-AI-path employment.

Claim ledger

Folded duplicate rows: none.

Estimator table

Nothing in this paper is estimated except one calibration input. Every headline magnitude is the output of a deterministic monthly recursion run at an assumed parameter bundle. The paper presents them as conditional scenarios, not predictions (p. 3), and none measures an effect in data.

Data integrity

Rung reached. Rung 1 (documentation gap) for the calibration, the survey and the printed constants. One item reaches rung 2 (verified error), in the explorer's implementation of the survey coding table, and no printed number can be shown to depend on it.

• Not deposited. The calibration code and the survey microdata, weights and coding are not deposited, so Tables 2 and 4 are unverifiable as published beyond their printed cells (analysis/code-and-data/data-findings.csv, items on the survey microdata and calibration code). The internal coherence checks pass: the steady-state finding-rate identity and the search-discount arithmetic.
• Observed-exposure anchor (0.14). Its construction from usage data is not deposited. A third-party recompute gives 0.1212, which moves extreme GDP from 32.41 to 32.20 and cognitive unemployment from 17.95 to 18.19 (work/logs/cd_calibration_inputs.log).
• Quits elasticity. It re-estimates as 0.5286 from the BLS originals of the cited series, under an inferred specification (work/logs/ts_battery.log; work/challenge/compensation-theory/s4_quits.log:5–14). The print rounds it "to 0.55". The code uses 6/11 = 0.5454, which rounds to 0.55 at two decimals; set at 0.55, four extreme-column cells move by at most 0.06. The kernel's own note derives 6/11 from the 0.06/0.11 occupational-move share instead, and the print does not say which rationale is meant.
• Table 3's income rows. They value cognitive labor two ways without saying so: labor income at wages paid, the labor share at the marginal product. The cognitive-firm term that reconciles them is negative, and Table A.1's "at most half a percent of GDP" holds in size, unsigned (code-and-data C2).
• Footnote 14's base. The footnote does not define it: 84.2 on mid-2026 employment, 87.0 on no-AI-path employment.
• Verified error (rung 2). For "less than 1 month" the explorer's quiz code computes 0.17 × 3/max(1, months) and gives μ = 0.51. Table B.1 prints "0.17×3/months, capped at 1", which gives 1 (p. 48; work/challenge/survey/s2_helpers.log:2, 34). Whether Table 4 used either cannot be checked.
• Flags with no rung.
  • The explorer runs capability and adoption answers below today's levels at the mid-2026 anchors (0.07 or 0 → 0.14; 0.05 → 0.10). The print does not say so.
  • The automation item's all-"alone" case is 0/0, and the print does not say how it is handled.
  • The solved logistic slope is capped at 3 a year and floored at 0, unstated; this does not bind in the scenarios.
  • Table 4 prints the net return's interquartile range as [7, 7].
  • Table 2 does not say whether it is weighted; Table 4's note says survey weights.
  • OEWS May 2025 gives a cognitive employment share of 0.560 on footnote 10's grouping. The kernel uses 0.6235 (CPS 2025) and has an OEWS May 2021 option of 0.614. This was not followed up (work/challenge/compensation-theory/s3_oews.log:48).

What the integrity tests covered.

• The authors' public implementation, against every printed cell of Tables 3, 5 and 6 and against the explorer's September record.
• The printed equation system, against the code (formula map, 37 rows).
• The one estimated input, against its public sources.
• The survey's printed quartiles, against the mappings, through synthetic respondents.

What they did not cover. The construction of the calibration targets from CPS, IPUMS and JOLTS; the usage-data anchor; the survey microdata. The paper is not cleared beyond these tests.

Findings

F1. Where the replacements of cognitive labor-force exits search sets the size of both unemployment headlines, of the extreme wage split and of the substantial cognitive-wage clause

bears on: "nearly one in five" (CL05), the mechanics sentence (CL10), the substantial rise (CL13), the extreme wage split (CL17), "slightly lower" cognitive wages (CL26), the cognitive-wage sentences (CL28) evidence: review/paper-REPORT.md (P9); review/code-and-data-REPORT.md (C19); analysis/code-and-data/assumption-ledger.csv (WA09); review/challenge-REPORT-scenarios.md elements 12, 19, 20, 27, 30, 34, 40; review/challenge-REPORT-new-rows.md elements 2 and 14; work/challenge/scenarios/s6_routing.log:1–17, 50–53; work/challenge/scenarios/s7_routing_spillovers.log:2–7; work/challenge/new-rows/s7_routing.log:3–15, 29–50; work/challenge/scenarios/s3_anchor.log:24, 26

Every quit from a cognitive job, including the 5/11 that are exits from the labor force, re-enters the pool as a cognitive-origin searcher. In the modest and substantial scenarios nobody is laid off. Across routings that keep the calibrated steady state:

• extreme cognitive unemployment is 13.85–17.62;
• substantial cognitive unemployment is 2.02–4.49;
• the extreme wage split and the sign of the cognitive-wage change both move;
• the substantial cognitive wage ends 1.03–1.45 percent above its no-AI path once a quarter or more of the replacements are redirected.

The premise is unanchored. Absorbed into the design-level finding on where the replacements of labor-force exits search (D2).

F2. The re-employment mapping's anchor contradicts the model's own job-finding rates; the eight-month ordering fails, and the survey's unemployment depends on how the answer is read

bears on: the abstract's survey sentence (CL06), the Introduction's survey sentence (CL07), the survey results paragraph (CL14), the eight-month ordering (CL15) evidence: review/paper-REPORT.md (P6); review/code-and-data-REPORT.md (C4); assumption ledger (WA02); review/claims-REPORT.md (L4); review/challenge-REPORT-survey.md elements 5–10, 27; work/challenge/survey/s4_core.log:22–50; work/challenge/survey/s7_extra.log:1–16

"Three months = 0.17" makes the normal-times answer stand for a 27.7-month cross-group wait in the model. Under every model-consistent reading, eight months is faster than the substantial scenario implies. Survey unemployment is at the modest level under cross-group readings and at the substantial level under the total-re-employment reading. Absorbed into the design-level finding on how the survey enters the model (D1).

F3. The own-hours capability measure that the paper prints puts the survey nearest the modest scenario

bears on: the abstract's survey sentence (CL06), the Introduction's survey sentence (CL07), the survey results paragraph (CL14) evidence: review/paper-REPORT.md (P7); review/code-and-data-REPORT.md (C5); assumption ledger (WA03); review/challenge-REPORT-survey.md elements 11–13; work/challenge/survey/s4_core.log:10–16

At m = 0.160 every outcome is nearest modest (GDP 3.70). The flip points (0.2175–0.2575) lie between the paper's two printed measures. The paper reports the 27-against-71 gap but does not run the alternative. Absorbed into the design-level finding on how the survey enters the model (D1).

F4. "The median respondent" is item-by-item medians, and the survey comparison is computed two ways

bears on: the abstract's survey sentence (CL06), the Introduction's survey sentence (CL07), the survey results paragraph (CL14) evidence: review/paper-REPORT.md (P2); review/challenge-REPORT-survey.md element 17; p. 49; Table 4 note (p. 36)

The model at the item medians, the Introduction's computation, is never printed. Table 4 prints medians of per-respondent runs over the complete cases, with four inputs fixed. Both land nearest substantial with the task-grid measure. The abstract's wording is looser than either computation, and p. 49 discloses that. Absorbed into the design-level finding on how the survey enters the model (D1).

F5. The calibration code, the survey microdata and the table-producing code were not deposited

bears on: the framework (CL01), the abstract's survey sentence (CL06), the survey results paragraph (CL14) evidence: 00_INTAKE.md (Deposit); analysis/code-and-data/data-findings.csv; review/code-and-data-REPORT.md (C9); work/logs/cd_calibration_inputs.log

Unverifiable as published:

• Table 4's per-respondent quantiles;
• whether Table 2 is weighted;
• the construction of the calibration targets, and of the 0.14 observed-exposure anchor (a third-party recompute gives 0.1212; extreme GDP 32.20, cognitive unemployment 18.19).

The survey rows are decided on the evidence that exists. The explorer's reproduction of every printed cell shows the tables are its outputs, not which code printed them.

F6. The compensation statements hold as potential compensation on every model-consistent capital closure; "nearly three times" is a gross-output multiple

bears on: the resources sentence (CL08), the textbook-compensation passage (CL16), footnote 14's "consume most of the gains" (CL18) evidence: review/paper-REPORT.md (P5); review/code-and-data-REPORT.md (C3, C12, C27); assumption ledger (WA01, WA12); review/claims-REPORT.md (L2); review/challenge-REPORT-compensation-theory.md elements 1–16, 37–40; work/challenge/compensation-theory/s2_kernel_battery.log:160–247; work/challenge/compensation-theory/s3_oews.log:47–50; work/challenge/compensation-theory/s5_share_variant.log:1–7; analysis/challenge-rerun-check.md part (i); work/audit/fix_comp.log sections B–D; work/verdict/grid_paper_values.log

Once the extra capital (1.96 times no-AI GDP) is charged, the multiple is 0.54–1.95. The rest stays ahead after the transfer on every closure except domestic financing of the 2030 build-up, at the scenario's own parameters and in every combination of the paper's parameter values. Paying the transfer leaves the gainers the whole net gain, which removes p. 37's "consume most of the gains" as a statement of what compensation leaves. The measured wage-bill share raises the multiple. The premise that the GDP gap is transferable is WHIMSICAL with a MATERIAL effect (the multiple and the 20-percent cushion), not a sign flip; the wage-bill-share premise is not FRAGILE, and is unresolved only for the transfer's size. Absorbed into the design-level finding on charging for the extreme scenario's capital (D3).

F7. The extreme unemployment figure is a corner of its own assumption bundle, and the paper varies neither search parameter

bears on: "nearly one in five" (CL05), the mechanics sentence (CL10) evidence: review/paper-REPORT.md (P8); review/code-and-data-REPORT.md (C6, C10); assumption ledger (WA06–WA08); review/challenge-REPORT-scenarios.md section (c), case against the authors 5; work/challenge/scenarios/s5_grid.log:89–116; work/challenge/scenarios/s4_variants.log:48–59

Where 17.9 sits in the grid. Over the paper's own values with calibrated ε and ξ, extreme cognitive unemployment spans 4.7–19.5 (median 9.4). It reaches 17.5 in 3.7 percent of those cells, and never when reinstatement is 0.5 (maximum 9.8 there). Over all the paper's values the main-effect shares are wage rigidity 0.56, reinstatement 0.14 and search discount 0.11 (s5_grid.log:96, 99, 107, 116).

One at a time.

• A search discount of 0.08 gives 14.3, and 0.17 gives 10.3.
• Reinstatement of 0.5 gives 8.8.
• Flexible wages give 2.6 (Table 6's own column).
• A slower posting speed raises the figure (0.25 / 0.10 → 18.5 / 19.5), so the bundle's two friction choices offset.

The robustness sections vary only ε and ξ.

Why the row does not narrow on this. The paper never attributes 17.9 to capability alone, and a scenario statement defines its bundle.

F8. Table 4's complete-case sample: the print does not bound it, random non-response does not move it, and the two survey tables are not reconciled

bears on: the abstract's survey sentence (CL06), the survey results paragraph (CL14) evidence: review/paper-REPORT.md (P2); review/code-and-data-REPORT.md (C22, C23); review/challenge-REPORT-survey.md elements 18, 19, 28 and section (c), case against the authors 4; work/challenge/survey/s4_core.log:74–75; work/challenge/survey/s5_synth.log:6–20; work/challenge/survey/s6_synth_seeds.log:6–19; analysis/challenge-rerun-check.md (survey qualification 6)

• The worst-case corners give GDP 1.8–31.7.
• Under random non-response the synthetic medians are 6.5–10.2, all nearest substantial.
• No run reproduces Table 4's GDP quartiles [3, 19].

Absorbed into the design-level finding on how the survey enters the model (D1).

F9. Several survey inputs do not change the nearest scenario; one premise is fragile and two coding choices are unstated

bears on: the abstract's survey sentence (CL06), the survey results paragraph (CL14) evidence: review/code-and-data-REPORT.md (C24–C27); assumption ledger (WA04, WA05); review/challenge-REPORT-survey.md elements 14–16, 20–22, 29; work/challenge/survey/s4_core.log:17–20, 55–73; work/challenge/survey/s7_extra.log:17–29; work/challenge/survey/s2_helpers.log:2, 34

These move the level but leave the survey nearest substantial: the fixed reinstatement and posting speed, the productivity code edge, and the present-tense reading.

Fragile or unstated:

• Adoption, coded as an instance share. The result flips below d = 0.1975.
• The explorer's lowest re-employment bin: 0.51 against the printed 1.
• The floors that run answers below today's levels at the anchors.

Absorbed into the design-level finding on how the survey enters the model (D1).

F10. The scenarios diverge mostly, not almost all, after the paper's date, and they share two of five of today's readings

bears on: the divergence sentence (CL09), the future-data sentence (CL23) evidence: review/paper-REPORT.md (P11); review/claims-REPORT.md (L5); assumption ledger (WA11); review/challenge-REPORT-scenarios.md elements 21–24; work/challenge/scenarios/s4_variants.log:12–27; work/challenge/scenarios/s3_anchor.log:23–27

How much divergence comes after the paper's date. For GDP and cognitive unemployment, 16.5–29.5 percent of the Jan-2030 gaps is in place by Jan 2028, the date "a year and a half from the mid-2026 anchor" names. So 70–84 percent comes after it. The small substantial-against-modest cognitive-wage gap is the exception: 34 percent of it is in place at mid-2026 and 54 percent by Jan 2028.

Which readings the scenarios share. They share the capability anchor (0.14) and the adoption anchor (0.10). They do not share:

• the productivity anchor (0.30 / 0.35 / 0.45);
• the automation share (0.5 / 0.75 / 0.9);
• reinstatement (0.5 / 0.25 / 0).

These sit at scenario values from 2024, although the paper reads today's use as about half automation on Claude.ai and three-quarters on the API (pp. 27–28). Starting the extreme automation share at today's API reading strengthens the timing slightly (5.2 → 4.6 percent of the GDP divergence by 2027).

Where the back-loading comes from. It is set by the assumed logistic paths. Each capability path also bridges a change of construct, from a usage anchor to endpoints on rated feasibility (p. 27).

F11. The small 2030 innovation effect is capped by the research-funding rule, whatever the ideas-block parameters

bears on: the innovation-channel sentence (CL12) evidence: review/paper-REPORT.md (P12); review/code-and-data-REPORT.md (C20); review/claims-REPORT.md (L6); review/challenge-REPORT-compensation-theory.md elements 23–25; work/challenge/compensation-theory/s2_kernel_battery.log:278–282; work/logs/th_ideas_fo_checks.log (b)

Research input is a fixed share of GDP on both paths (Equation 22). So with a nonnegative ideas gap and λ ≤ 1, the 2030 ideas gap is at most g times the cumulated GDP gap: 0.639–0.661 against the simulated 0.605 in the extreme scenario, and 0.211–0.216 against 0.198 in the substantial. The paper derives this smallness itself (pp. 16–17) and calls the effect a lower bound. What does not hold is the label "surprising finding" on p. 4.

F28. The Introduction's "in each scenario ... cognitive wages grow more slowly than they would without AI" fails for the modest scenario, which the paper's own results print

bears on: the cognitive-wage sentences (CL28) evidence: review/challenge-REPORT-new-rows.md elements 1–14; work/challenge/new-rows/s3_scan.log:10, 26, 29–31, 44, 49, 190–192, 228; work/challenge/new-rows/s4_variants.log:4, 21–32; work/challenge/new-rows/s6_relative.log:2–4; work/challenge/new-rows/s7_routing.log:18–39; work/audit/aud_fig3_wages.log:27–32; Table 3 (p. 31); Figure 3 (p. 33, work/challenge/new-rows/pages/p33_cogwage_crop-33.png); pp. 4, 7, 32, 38

The modest scenario. As specified, the modest cognitive wage is above its no-AI path in every month from Jan 2024, +0.41 in 2030, and grows faster than without AI in 59 of 61 twelve-month windows. The two exceptions start in Jan and Feb 2024 and come from the AI paths' disclosed start (p. 40). It stays above the path through 2035 (+1.44) and under every routing of labor-force exits. The wage that would clear the cognitive group is above the no-AI path from Jan 2025 too. The paper prints the sign itself: Table 3 (+0.4), Figure 3 ("Modest 0.4" above the no-AI line) and p. 32 ("the cognitive wage up 0.4 percent"). The conclusion states the narrower claim (p. 38). The overstatement is the Introduction's, and the same comparator appears in the p. 7 overview ("relative to the paths they would have taken without AI").

What holds. As specified, the substantial and extreme cognitive wages are below their no-AI paths in every month and grow more slowly in every window. Wages elsewhere grow faster and employment shifts toward the other occupations in all three scenarios. Read as the displacement channel, the p. 3 sentence holds everywhere: the cognitive wage is below the common wage in every month of every scenario.

What depends on a premise. The sign of the cognitive wage against its no-AI path is not structural: it depends on wage rigidity (modest −0.24 at ξ 0; Table 6 prints positive cognitive wages in three of its seven columns). The substantial clause rests on the exits premise (D2).

F12. The substantial scenario's cognitive unemployment is 3.06 percent in mid-2026, not 2.9; the printed 2.9 is the no-AI level

bears on: the substantial rise (CL13) evidence: review/challenge-REPORT-scenarios.md elements 30–31; work/challenge/scenarios/s4_variants.log:5; Figure 4 and its note (p. 34)

As specified, the rate rises from 3.06 in mid-2026 to 4.53 in Jan 2030, an increase of 48 percent. The 2030 rate is 59 percent above the no-AI 2.85. The p. 34 sentence dates the no-AI level to mid-2026; Figure 4's note gives 2.9 as the "normal" level. The size of the rise also rests on the exits routing (D2).

F24. Footnote 14's 84 percent reproduces on one of two baselines, its example is so-so only in p. 14's relative sense, and "consume most of the gains" misreads what compensation leaves

bears on: the so-so example (CL18) evidence: review/code-and-data-REPORT.md (C7); review/paper-REPORT.md (P4); review/challenge-REPORT-compensation-theory.md elements 17–21, 39–40; work/challenge/compensation-theory/s2_kernel_battery.log:241–247; analysis/challenge-rerun-check.md part (i) and qualification 4; footnote 14 (p. 37); p. 14

The figure is 84.2 on mid-2026 employment and 87.0 on no-AI-path employment, and the paper uses both baselines without saying so. The example keeps a 57 percent per-instance gain. A setting that is so-so in p. 14's sense gives transfers of 173 percent and 69 percent of the gain (ε 1 and 3). After the transfer the gainers keep the whole 7.21-point net gain, and the transfer is 46 percent of their gross gain; "would consume most of the gains" holds only as a comparison of sizes. The identity is part of the design-level finding on charging for capital (D3).

F13. No exercise shows that coming data will tell the scenarios apart

bears on: the future-data sentence (CL23) evidence: review/code-and-data-REPORT.md (C15); analysis/code-and-data/quantitative-model-diagnostics-REPORT.md (QM7, QM8); review/paper-REPORT.md (P11); review/challenge-REPORT-survey.md elements 23–26; work/challenge/survey/s7_extra.log:30–41; work/logs/qm_untargeted.log

No moment tests the fit. The steady state reproduces its labor-market targets by construction. Inside the data window, the mid-2026 values the scenarios imply (GDP 0.55 / 0.83 / 1.21 above a no-AI path; cognitive unemployment 2.89 / 3.06 / 3.33) have no observed counterpart that isolates AI.

Several scenario parameters have no anchor. Reinstatement, the search discount on the path and posting speed have none. ε = 3 rests on half a long-run wealth elasticity applied over four years.

The obstacle is attribution, not size. By Jan 2028 the extreme scenario's gaps from the others are large: cognitive unemployment 6.07 against 2.91 and 3.38, and GDP gaps of 4.0–5.6 points. The modest–substantial gaps are small (GDP 1.6 points, cognitive unemployment 0.47). Telling the scenarios apart would require attributing those gaps to AI. Neither the Jan-2028 share of the divergence nor the capital-supply passage bears on whether data can discriminate.

F21. The benchmark sentences hold as levels, or as a pace from 2024, not on the cited works' horizons and not for every work cited

bears on: the benchmark sentences (CL25) evidence: review/interim-closeout-REPORT.md (IC2–IC4); work/interim/benchmark_scenario_side.log:3, 6, 9–13; work/audit/benchmark_lit_REPORT.md; work/audit/benchmark_lit.log (L014–L145); review/extracts/Acemoglu_2024_SimpleMacroeconomicsAI_NBERw32487.txt (PDF pp. 2, 6, 45); review/extracts/Briggs_2023_LargeEffectsAIGrowth.txt (printed pp. 1, 13, 14); review/extracts/Karabarbounis_2014_GlobalDeclineLaborShare_NBERw19136.txt (PDF p. 3) and its Figure 2 (work/interim/kn_p38-38.png); review/extracts/Filippucci_2024_MiracleOrMyth_OECDAIP29.txt, Filippucci_2025_G7ProductivityGainsAI_OECDAIP41.txt, Filippucci_2026_AggregateGainsAISectoral_AEAPP_IDEASrecord.txt, McKinsey_2023_EconomicPotentialGenAI.txt, Baily_2023_MachinesOfMind.txt, Aghion_2024_AIandGrowthWhereDoWeStand.txt, Arnon_2025_PWBM_ProjectedImpactGenAI.txt; review/challenge-REPORT-scenarios.md elements 36–37; work/challenge/scenarios/s8_benchmarks.log:1–3; pp. 3, 33

The modest scenario.

• Acemoglu. The modest scenario's 2030 GDP gap (1.61) is slightly above his upper, investment-boom range (1.4–1.56 percent over ten years), and its TFP gap (0.71) slightly above his 0.66 bound. Against his baseline range (0.93–1.16) it is 1.4 to 1.7 times as large. His figures accrue over ten years and the scenario's over six from 2024, so GDP accrues 1.7 to 2.3 times, and TFP 1.8 times, as fast.
• The OECD figures (Filippucci, Gal and Schief 2024; Filippucci, Gal, Laengle and Schief 2025; the AEA Papers and Proceedings abstract). All of them lie above the modest scenario: annual US TFP gains of 0.24–0.62 points are 2.0 to 5.2 times its pace, labor-productivity gains of 0.36–0.93 points 1.3 to 3.3 times, the 2025 paper's US range (0.41–1.28) 1.5 to 4.6 times, and the abstract's "up to 0.9" of TFP 7.6 times (benchmark_lit.log, L032–L035, L049, L056). That is the same order of magnitude, but the scenario sits between Acemoglu and the OECD's lowest figure, not inside the OECD range. The 2025 paper's medium and rapid adoption cases are closer to the substantial scenario (0.68 and 0.88 times its pace).

The substantial scenario.

• Briggs and Kodnani. The substantial scenario adds 1.45 points a year to output per worker from 2024, matching their figure of just under 1.5 points a year. They put that pace in a ten-year window that opens once about half of businesses have adopted, at a date they call hard to predict, so they do not forecast 8.3 percent by 2030. From the mid-2026 anchor the scenario's pace is 2.23 points a year, about 1.5 times theirs.
• The works said to give "comparable magnitudes." Aghion and Bunel's figures are 0.55 to 0.90 times the substantial pace on labor productivity and 1.33 times on TFP. Baily, Brynjolfsson and Korinek's illustrative 18 percent over ten years is 1.15 times; the forecast they endorse as reasonable, 7 percent of global GDP over ten years, is 0.47 times. McKinsey's generative-AI contribution (0.1–0.6 points a year globally, 0.3–0.7 in the US) is 0.07 to 0.48 times; its all-automation range brackets the scenario (0.34 to 2.34 times). The Penn Wharton brief's yearly TFP contributions cumulate to a GDP gain of 0.60 percent by 2030, 0.073 times the substantial scenario's 8.29 and below the modest scenario's 1.61; at its own 2035 horizon it reaches 1.55 percent (benchmark_lit.log, L073–L075, L088, L101–L102, L109–L114). The brief obtained carries a different title from the one cited.

The extreme scenario. At 32.4 percent of GDP by 2030 it is at least four times every AI-specific figure cited, put on the same 2030 horizon, and 3.1 times Baily et al.'s illustrative path. Against McKinsey's all-automation upper end it is only 1.29 to 1.46 times. On McKinsey's generative-AI figure the substantial comparison fails and the extreme one holds; on its all-automation range the reverse. The paper does not say which it means (benchmark_lit.log, L117–L143).

The labor share. "Four points in four years" rounds 3.43 points in 3.5 years, and p. 33's "60 percent of income in 2026" is the 2024 base. That is slightly less than the fall of roughly 4 to 5 points to 2011 in the US corporate labor share of Karabarbounis and Neiman's Figure 2, read by eye. But their data end in 2012, about three decades after 1980, not four; they measure the corporate sector, not the aggregate share the model reports; and their text's only figure is global, 5 points over 35 years.

How the comparisons were built. Each work's annual rate is put on the scenarios' horizon as six years of compounding from 2024, a construction no source reports; it treats the works' ten-year or seventeen-year averages as uniform, and both the works' adoption paths and the scenarios' gaps are back-loaded. McKinsey's figures are productivity growth rates that assume displaced hours are redeployed, not GDP levels.

The published versions of Acemoglu and of Karabarbounis and Neiman, and the full text of the AEA Papers and Proceedings article, could not be obtained, operator access included, and are unverified here (Status). The figures above for those works are the working papers' and the abstract's; a figure changed in a published version would change these comparisons.

F14. Footnote 3's approximate wage threshold is less than half the exact one at the extreme parameters

bears on: footnote 3 (CL27) evidence: review/code-and-data-REPORT.md (C16); analysis/code-and-data/theory-diagnostics-REPORT.md (TH1); review/challenge-REPORT-compensation-theory.md elements 34–36; work/challenge/compensation-theory/s2_kernel_battery.log:252–260; page image work/challenge/compensation-theory/pages/p-14.png

The first-order threshold is 0.96 against an exact 2.10 (extreme) and 1.30 against 1.76 (substantial). At ε = 1 in the extreme scenario, the first-order wage rises 0.55 percent while the exact wage falls 14.3 and Table 5 prints −9.2. The share clause is 53 percent first order against 24 exact.

The "≈" on the threshold means the "if and only if" holds for the exact value. No printed table rests on the approximation.

F15. "Slightly lower" cognitive wages are a real fall of about 6 percent, and no-AI wage growth is 1.67, not 2, percent

bears on: "slightly lower" cognitive wages (CL26) evidence: review/paper-REPORT.md (P4); review/claims-REPORT.md (L8); review/challenge-REPORT-scenarios.md elements 38–40; work/challenge/scenarios/s4_variants.log:8–9; work/challenge/scenarios/s7_routing_spillovers.log:2–7; Table 1 (p. 23)

The fall from mid-2026 to Jan 2030 is 5.96 percent at the model's no-AI wage growth of 1/60 a year, and 4.85 at the sentence's 2 percent (2 percent is GDP growth). Its sign rests on the exits routing (D2).

F16. The authors' implementation reproduces every printed cell, solves what it states, and matches Proposition 1

bears on: the framework (CL01), Proposition 1 (CL20) evidence: review/code-and-data-REPORT.md (C1, C14, C17); work/logs/cd_reproduce.log; work/logs/cd_record_compare.log; work/audit/rerun/logs__cd_econark_pytest.log; work/challenge/scenarios/s3_anchor.log:22; work/challenge/scenarios/s5_grid.log:117; work/challenge/compensation-theory/s2_kernel_battery.log:157, 263–275

• Tables 3, 5 and 6 reproduce to the printed rounding, with every minus sign checked on the page image.
• The current chunk matches the September record to 3.2e-14.
• 7,776 of 7,776 grid runs solve.
• Equations 14–18 equal the kernel's target economy to 4.4e-16.
• The match uses the model's own no-AI ideas growth of 1/60; at the printed 1.67, three cells miss by up to 0.06.

F17. Extreme GDP and the labor share are outputs of their bundle and move mainly with the capital-supply elasticity and reinstatement; ε = 3 has a mismatched anchor

bears on: extreme GDP and growth (CL03), the labor share (CL04) evidence: review/code-and-data-REPORT.md (C11, C15); analysis/code-and-data/quantitative-model-diagnostics-REPORT.md (QM2, QM8); review/challenge-REPORT-scenarios.md element 6; work/challenge/scenarios/s5_grid.log:73–106; Table 5 (p. 37)

Over the paper's own values:

• extreme GDP spans 15.7–52.9 (main-effect shares: ε 0.41, reinstatement 0.28);
• the extreme labor share spans 40.5–58.5, so "falls by a quarter" needs reinstatement 0 and automation near 0.9.

ε = 3 is half a long-run wealth elasticity used over four years, and it is the largest GDP driver. Table 5 discloses its effect. The substantial scenario sits near the middle of its grid (GDP 8.3 against a median of 7.8).

F18. Table 3's income rows use two wage concepts without saying so, and the cognitive-firm term is negative

bears on: the labor share (CL04), total labor income (CL19) evidence: review/paper-REPORT.md (P3); review/code-and-data-REPORT.md (C2); review/challenge-REPORT-compensation-theory.md elements 26–27, 41; work/challenge/compensation-theory/s2_kernel_battery.log:286–289; Table 3 and its note (p. 31); Table A.1 (p. 43)

In the extreme column, labor income (+0.5) plus capital income (+81.4) exceed GDP (+32.4) by the cognitive firms' loss. The share and GDP rows imply labor income of −0.2, at the marginal product. The firm term is −0.31 percent of GDP in 2030 and negative in 901 of 982 non-zero months (largest 0.47). That is inside Table A.1's unsigned bound. No row is wrong by its own definition.

F19. The modest scenario's sentence survives every alternative value tried

bears on: the modest scenario (CL02) evidence: work/logs/claims_overturn.log; work/challenge/scenarios/s4_variants.log:34–46

Growth rises +0.33 to +0.47 points and unemployment changes −0.12 to +0.13 points, one at a time; the worst combination gives +0.29. A combination that also gives the modest scenario the extreme automation and reinstatement reaches +0.73, but that replaces the scenario's own bundle and does not test the sentence.

F20. The within-task elasticity is load-bearing for the labor share and unanchored, and it moves the share in the direction that strengthens the fall

bears on: the labor share (CL04) evidence: review/paper-REPORT.md (P10); review/code-and-data-REPORT.md (C18); assumption ledger (WA10); work/logs/th_nested_sigma.log

An AI-performed and a human-performed instance of the same task substitute at σ = 0.5 by assumption (p. 8). The cited evidence concerns substitution across tasks or occupation groups. In a nested variant (target economy, extreme 2030, ε 3), the labor share is 44.6 at a within-task elasticity of 0.5, 42.4 at 1 and 37.4 at 3. The share leaves "45 percent" once the elasticity exceeds about 0.65, downward. The unemployment consequences were not simulated.

F25. Capital supply and wage rigidity move outcomes as printed; worker-side responses would reshape the flexible-wage column

bears on: capital supply and wage rigidity (CL11) evidence: review/code-and-data-REPORT.md (C1, C19); review/paper-REPORT.md (P9, second paragraph); work/logs/th_mobility.log; Tables 5 and 6 (pp. 37–38)

With flexible wages, the cognitive wage is 42.2 percent below its no-AI path and the other wage 70.1 percent above, yet cognitive employment falls only 1.3 percent: nothing in the model lets workers move toward the higher wage. With illustrative responses (quits responding to the wage gap; directed search), the cognitive wages at ξ 0 / 0.5 / 0.9 are −23.8 / −2.5 / +5.0 against the printed −42.2 / −11.5 / +2.8. The direction holds; the strengths are not calibrated.

F22. Both ideas-channel appendix claims hold

bears on: the ideas appendix claims (CL21) evidence: review/code-and-data-REPORT.md (C13); analysis/code-and-data/quantitative-model-diagnostics-REPORT.md (QM4, QM5); work/logs/qm_ideas.log; work/challenge/compensation-theory/s2_kernel_battery.log:282–283

The closed-form gap is inside 0.02 points with 0.0004–0.0006 to spare. Letting ideas augment capital moves 2030 cells by at most 0.9 points but 2035 extreme GDP by +17.6.

F23. The quits elasticity reproduces under an inferred specification and moves no headline

bears on: the quits elasticity (CL22) evidence: review/code-and-data-REPORT.md (C21); analysis/code-and-data/time-series-diagnostics-REPORT.md (TS1–TS5); work/challenge/compensation-theory/s4_quits.log:5–14; work/challenge/compensation-theory/s2_kernel_battery.log:292–298

The elasticity is 0.52–0.54 on the hazard and 0.63–0.64 on the probability. Month to month the response is small, and the formal cointegration tests cannot decide. Across 0 to 1, extreme cognitive unemployment moves within 17.6–18.2. The code's 6/11 differs from the printed 0.55 by 0.0046.

F29. Capital's return and share rise, and the labor share falls, in every month of every scenario

bears on: capital's return and share (CL29) evidence: review/challenge-REPORT-new-rows.md elements 15–20; work/challenge/new-rows/s3_scan.log:36–42, 90–96, 144–150; work/challenge/new-rows/s4_variants.log:6, 11, 16, 35–46; work/challenge/new-rows/s7_routing.log:18–50; work/audit/aud_fig3_wages.log:34–36; Table 3 (p. 31); pp. 9, 40

The rental rate, the net return, the capital stock and the capital share are above their no-AI values in all 73 months from Jan 2024 to Jan 2030, strictly, in every scenario, at every finite capital-supply elasticity the paper prints and under every routing of labor-force exits. The labor share falls in every month after Feb 2024, and the GDP gap exceeds the average-wage gap from Jan 2025. At a pegged rental rate the return does not rise, a boundary the paper states; the scenarios use ε = 3.

F26. The modest and substantial reallocations are inside documented US occupational shifts in size, at the fast edge in speed

bears on: "absorbed historically" (CL24) evidence: review/interim-closeout-REPORT.md (IC1); work/interim/reallocation_vs_history.log:1, 3, 5, 9–21; review/extracts/Dimitri_2005_20thCenturyTransformationAgriculture.txt (PDF p. 5); review/extracts/Autor_2019_WorkOfThePast_NBERw25588.txt (printed p. 7 and footnote 6); review/extracts/Atkinson_2017_FalseAlarmismOccupationalChange.txt (printed pp. 13–14, 17); work/challenge/scenarios/s8_benchmarks.log:5; work/challenge/scenarios/s6_routing.log:4–9; p. 35

The substantial scenario lowers the cognitive share of employment 2.01 points in 3.5 years (0.57 a year), and a net 1.52 percent of employment leaves the cognitive group; the modest scenario lowers it 0.26 points. Documented broad-group shifts, at their decade or multi-decade average paces, move a share 1.2 to 2.3 points in 3.5 years, and proportionally the substantial shift is slower than all of them (0.94 percent a year, against 1.11 for middle-skill hours and 2.0–5.6 for agriculture). In points a year only agriculture's 1900–30 decline (0.65) is faster. Every comparator is a 10-to-45-year average on a different base (workforce, hours, employment), and no 3.5-year window was obtained. Unemployment in 2030, 4.57 overall and 4.53 cognitive, is inside past peaks of about 10 and 8 percent, and every routing lowers it. The comparison can fail: the extreme scenario's 8.93 points (2.55 a year) is far outside. Two comparisons with Atkinson and Wu's detailed-occupation measures (churn, and jobs in shrinking occupations) are not counted: a two-group measure understates the detailed one, so their favourable results say little.

F27. Minor printed points with no bearing on a ledger row

bears on: no ledger row evidence: review/paper-REPORT.md (P4); review/code-and-data-REPORT.md (C8); work/logs/cd_reproduce.log; Table 4 (p. 36); Table A.2 (p. 45)

• ψ is "held constant" (p. 11) and is constant in the code; Table A.2's logistic for ψ is unused generality.
• Table 4's net-return interquartile range [7, 7] hides dispersion under integer rounding (the capital stock's is [6, 27]).
• The t0 GDP gap is 0.245 / 0.126 / 0.036 percent, so "at most a quarter of a percent" (p. 40) holds, narrowly.

Limits on the review

• Unobtainable versions. The earlier draft acknowledged on p. 1 is unobservable, and the body text of CEPR DP21939 is unverified here (paywalled). The published Economic Policy (2025) article of Acemoglu, the QJE (2014) article of Karabarbounis and Neiman and the full text of Filippucci, Gal and Schief's AEA Papers and Proceedings article, all cited by the benchmark sentences, could not be obtained after every acquisition tier, operator access included; the NBER working papers and the AEA article's abstract record were read instead (00_INTAKE.md, acquisition notes). The Penn Wharton brief read carries a different title from the one cited; the operator confirmed it as the cited item.
• Where replacements search. Nothing in the packet measures where the replacements of labor-force exits search in 2026–30. The routing family bounds the unemployment headlines and the cognitive-wage signs; it does not measure them.
• Variants not built or not swept.
  • A heterogeneous-pay version of the model (employment share 0.624 but wage-bill share 0.69–0.75) was not built: the kernel has no such switch, and the challenge's share variant moves both shares together. It would settle the transfer's size at the measured wage-bill share.
  • The unemployment consequences of a within-task elasticity above 0.5 were not simulated; the nested variant covers the target economy only.
  • The worker-side response strengths are illustrative, not calibrated.
  • The challenge's synthetic-respondent marginals were not swept, in particular their lower tails.
  • The exits routing was not combined with the rigidity or capital-elasticity variants; each premise was changed one at a time.
• The benchmark constructions. Putting each work's annual rate on the scenarios' 2030 horizon is the review's construction, approximate in both directions (work/audit/benchmark_lit_REPORT.md, conversion rule). The "best benchmarked against" works the extreme sentence names (Korinek 2023; Korinek and Suh 2024; Davidson 2023; Erdil et al. 2025; Kokotajlo et al. 2025) were not read: the sentence names them as framing, not as figures.
• What the challenge did not rebuild stands as the stage reports give it: the September drift check, the third-party rebuild's test suite, the capital-augmenting ideas variant, the nested within-task variant, the worker-side variants, the cumulative 2026–30 compensation figures (a supply-price gain of 0.087 of 2030 no-AI GDP against transfers of 0.129, a ratio, not a shortfall), the Figure 3 readouts at axis dates beyond those the challenge of the cognitive-wage sentences rebuilt, and the review's joint grid itself. That grid's compensation shares were recomputed from its saved outputs with the compensation identity (work/audit/fix_comp.log section C; work/verdict/grid_paper_values.log), and its paper-values GDP range matches the scenarios challenge's own grid.
• Unchecked links.
  • Whether the explorer's quiz code is the code that built Tables 2 and 4.
  • The representativeness of the online panel beyond what Appendix B prints.
  • The OEWS-versus-CPS employment-share gap (0.560 against 0.6235), which was not followed up.
  • Whether the Introduction's wage sentence dates from an earlier calibration: only one version of the paper is observable.

Counts

Rows whose holds moved from the claims stage's (3), all toward not established.

• The extreme wage split, "the cognitive wage is 11.5 percent below its no-AI path while the wage in all other occupations is 34 percent above it" (CL17): holds → narrower. The stated sizes rest on the same exits-routing premise that narrows the unemployment rows (work/challenge/scenarios/s7_routing_spillovers.log:2–7; the same figures on the second routing copy, work/challenge/new-rows/s7_routing.log:44–50).
• The future-data sentence, "those data will tell us which scenario we are in" (CL23): narrower → not established. Its central assertion has no direct test, and the challenge showed the two points the claims stage offered do not bear on it (review/challenge-REPORT-survey.md, elements 24–25).
• The benchmark sentences (pp. 3, 33; CL25): holds → narrower. The cited works, read after the claims stage, put their figures on horizons the sentences leave out (review/interim-closeout-REPORT.md, IC2–IC4), and several further works the same sentences cite give figures far from the stated magnitudes (work/audit/benchmark_lit_REPORT.md; work/audit/benchmark_lit.log).

No row moved toward holds. The count is one-sided because none of the three rests on new evidence against the model the claims stage tested: one applies the claims stage's own exits premise to every row it carries, one applies the rule that a claim whose only direct test is unrun is not established, and one rests on the cited works themselves, read for the first time, and narrows framing sentences only.

A row whose holds matches the claims stage's on other grounds. The so-so example (footnote 14, p. 37; CL18) is narrower, as the claims stage had it, but not for its reasons: 84 percent reproduces as 84.2 on mid-2026 employment, and p. 14 defines so-so automation in relative terms, so neither ground stands (work/challenge/compensation-theory/s2_kernel_battery.log:241–247). The row is narrower because "compensating workers would consume most of the gains" holds only as a comparison of sizes: paying the transfer leaves the gainers the whole net gain (analysis/challenge-rerun-check.md, part (i); review/challenge-REPORT-compensation-theory.md, element 40).

What moved relative to the previous synthesis (git show fc2fe57:SYNTHESIS.md: 27 rows, 2 not established / 13 narrower / 12 hold; this one: 29 rows, 2 / 15 / 12).

• Two rows added. The cognitive-wage sentences (pp. 3–4; CL28; narrower) and capital's return and share (pp. 3–4; CL29; holds), which the audit's completeness check found the ledger lacked, were written by the claims stage and tested by a fresh challenge (review/challenge-REPORT-new-rows.md). The cognitive-wage row is narrower on the modest scenario's printed sign and on the exits premise (findings F28 and D2); the capital row holds (F29).
• The so-so example (CL18): holds → narrower. The previous synthesis read "consume most of the gains" as a comparison of sizes only and recorded the identity under what survives. Read as also saying what compensation leaves, the sentence is contradicted by the identity, and a reader taking it so would misjudge the cost of compensation; that is more than a nuance. The evidence is unchanged: s2_kernel_battery.log:241–243 and the identity.
• No other row's holds moved.

Re-decided on the routed corrections (CHANGELOG.md, entry "29 Oct 2026 — audit-review against review 0.9.2").

• The added rows (correction V1). They bear on two new ledger rows and the bottom line. The cognitive-wage row stands narrower, as the claims stage proposed, with its kernel narrowed further to what the challenge's re-run evidence supports (the substantial clause under the exits premise; the p. 3 sentence as a channel). The capital row holds. The design-level finding on where replacements search gains the substantial cognitive-wage clause and a table row; two findings are added (F28, F29); the bottom line gains the wage sentence and the capital sentence.
• The benchmark works (V2). They bear on the benchmark row. It stays narrower: the OECD, McKinsey, Baily et al., Aghion–Bunel and Penn Wharton comparisons are added to its kernel and not-established text, its defence is restated, and its status names the AEA Papers and Proceedings article as unverified here. The benchmark finding (F21), the bottom line's benchmark sentence, Status and the limits change with it.
• The compensation correction's consequences (V3). They bear on the design-level finding on charging for capital, the resources and textbook-compensation rows and the so-so example. The finding carries the grid result (the foreign-owned rest ahead in every combination of the paper's parameter values, +1.36 to +15.48; "more than 20 percent ahead" in 89 / 52 / 29 / 0 percent of them on the GDP, net-product, replacement and foreign-owned bases) and states the premise labels: the transferable-GDP-gap premise stays WHIMSICAL with a MATERIAL effect, not a sign flip; the wage-bill-share premise does not stay FRAGILE. The resources and textbook-compensation rows stay narrower, their kernels carrying the grid. The so-so example moves, as above. The textbook-compensation row's not-established text adds the transfer's size at the measured wage-bill share, which no construction built settles.
• TEXT corrections already in the previous synthesis (CHANGELOG.md, items T2, T4–T8, T10) are kept: no re-run notes in Status; no superseded stage figures; the third-party rebuild's 194 tests; the two verbatim quotations; the resources row's defence verbatim; the grid result for the foreign-owned closure; the log paths for the no-AI unemployment rate and the overall rate under every routing.

Changed outside the routed corrections.

• The grid's description. The review's 2,592-cell extreme grid adds a steeper-logistic factor the review introduced. The synthesis reports its 1,296 paper-values cells, recomputed here (work/verdict/grid_paper_values.py, work/verdict/grid_paper_values.log); their GDP-basis range, 15.66–52.87, matches the scenarios challenge's independent grid (work/challenge/scenarios/s5_grid.log:80).
• The extreme-unemployment finding's main-effect shares are restated to the paper's values (wage rigidity 0.56, reinstatement 0.14, search discount 0.11; s5_grid.log:107).
• The routing family's AI-off check for the cognitive wage (at most 0.016 points) and the modest cognitive wage under the routings (+0.42 to +0.60, the share rule that fails the AI-off check excluded) are added from the second routing copy (work/challenge/new-rows/s7_routing.log:9–28).
• Findings order. The footnote-14 finding moves up with its row; the capital-supply finding sits with the headline rows that hold, ahead of the supporting ones.
• Kept from the interim close-out: its two comparisons with Atkinson and Wu's detailed-occupation measures stay uncounted, because a two-group measure understates the detailed one, so a favourable result says little (review/interim-closeout-REPORT.md, IC1).

Status changes without a holds move (against the claims stage's provisional status).

• "Nearly one in five" (CL05) and the substantial rise (CL13) went from INTERIM to settled: the challenge ran the pending routing rebuild.
• The textbook-compensation passage (CL16) went from INTERIM to settled: the challenge measured the wage-bill share.
• "Absorbed historically" (CL24) went from INTERIM to settled: the interim close-out ran the historical comparison.
• The benchmark sentences (CL25) went from INTERIM to unverified-here, decided on the evidence that exists.
• No row is INTERIM (Status).

Challenge rulings by direction. Element rulings, summed from the four reports' count tables (scenarios, survey, compensation-theory, new rows):

Claim-level rulings:

• STANDS, 8: the framework, the modest scenario, the labor share, capital supply and rigidity, Proposition 1, the ideas appendix claims, the quits elasticity, and "absorbed historically".
• STANDS WITH CHANGES, 21, including both added rows.
• WRONG, 0.

All ten WRONG rulings fell on elements the review had counted against the authors, eight of them in the compensation closures. The review's errors ran against the paper, and this synthesis carries the corrections.

Mechanical re-run. Every value cited by a WRONG or STANDS WITH CHANGES ruling of the three original groups re-ran identically, and every page matched (analysis/challenge-rerun-check.md). The new-rows group's four load-bearing scripts re-ran identically and left the files they rewrite unchanged (work/verdict/rerun/s3_scan.log, s4_variants.log, s6_relative.log, s7_routing.log), and its cited pages were opened: Table 3 (p. 31), the p. 4 paragraph and Figure 3's cognitive-wage panel (p. 33) as images; pp. 3, 7, 9, 32, 38 and 40 in the text. Qualifications carried where they bear:

• the unprinted base of "more than 20 percent ahead";
• the hand-derived 46 percent of the gainers' gross gain;
• the 0.03–0.05-point AI-off offset of the routing rules for unemployment, and at most 0.016 points for the cognitive wage;
• one survey citation line whose median was 9.08 rather than the matched run's;
• the Table B.1 multiples;
• the new-rows challenge's own notes that the audit's capital scan tested "not below" rather than "above" (the strict test holds) and that its "Tested" on the displacement reading was argued, then computed by the challenge.

No ruling was set aside.