[MaxGhenis]

The Budget Lab revised Iselin and Nunn, "How potential AI futures would play out in the current tax system", on 9 October 2026 (AI-Fiscal v1.0.1). Our replication pinned the July Table 1, so it no longer matched. This PR moves it to the revised paper, reruns every scenario on the published runtime, and regenerates the site, memo and paper.

What the Budget Lab changed

From the revised report's erratum and AI-Fiscal's v1.0.1 changelog:

1. 2030 labor shares. Both inputs now come from the Economists column of Karger et al. (2026): Table 19 for GDP growth, Table 39 for labor shares. In July the labor shares came from the pooled Total column. Moderate 53.8% → 54.0%, Rapid 51.3% → 52.0%; Slow is 55.0% in both. In their Table 1, capital growth goes Moderate 7.54% → 7.08% and Rapid 17.28% → 15.60%; labor growth Moderate 0.41% → 0.78% and Rapid −0.94% → +0.41%. GDP growth and λ are unchanged.
2. CBO FY2030 ratios. The CIT/GDP ratio moves from the rounded 1.3% to 1.2766%, so their corporate anchor goes from $486.1B to $477.3B and their macro CIT wedge falls 1.8% in every cell. Total revenue is now $6,595.1B (was $6,618.2B). This changes their side of our bridge only; our microsimulation does not use it.

Their headline, Rapid / expansive with the corporate wedge, goes from $216.0B (3.3%) to $246.3B (3.7%). The July wording ("roughly twice as large", "$216 billion, a 3.3 percent increase", labor shares "55.0%, 53.8%, or 51.3%", ATR "21% … to 34%") is confirmed against the Wayback Machine's 21 July 2026 copy of the report.

Why ours moved

Only through (1). Same runtime, same data: policyengine.py 5.0.1, policyengine-us 1.764.6, build p, Head Start excluded, runtime fingerprint dc91de12d0ddbc57. The 2030 baseline row is byte-identical to the July-inputs run.

One more input change, of no practical size. The CBO growth factor is now (1.022)(1.018)^4 = 1.0975947, the path the report states and AI-Fiscal commits, in place of a four-decimal 1.0976 recovered from the July table. With 1.0976 the revised Rapid capital growth is 15.5949%, which rounds to 15.59% rather than the published 15.60%. With the exact factor and θ0_L = 0.555, our g_Y, g_K, g_L and λ match AI-Fiscal's committed per-cell parameters to 1e-16 in both releases. Slow and the shares-fixed counterfactuals, which the labor-share change does not touch, move by at most $0.04B.

Before (July inputs) → after (revised inputs), 2030

Cell	Labor share	Revenue $B	Payroll $B	SPM poverty Δ pp	Net Gini Δ
Slow / shares fixed	fixed	+47.2 → +47.3	+10.1 → +10.1	-0.14 → -0.14	+0.0002 → +0.0002
Slow / compressive	55.0% → 55.0%	+3.3 → +3.3	-1.1 → -1.1	-0.12 → -0.12	-0.0003 → -0.0003
Slow / proportional	55.0% → 55.0%	+6.5 → +6.5	-5.4 → -5.4	+0.04 → +0.04	+0.0007 → +0.0007
Slow / expansive	55.0% → 55.0%	+9.8 → +9.9	-9.7 → -9.7	+0.20 → +0.20	+0.0018 → +0.0018
Moderate / shares fixed	fixed	+286.8 → +286.8	+61.1 → +61.1	-0.67 → -0.67	+0.0013 → +0.0013
Moderate / compressive	53.8% → 54.0%	+124.5 → +141.2	+33.0 → +39.5	-0.95 → -1.02	-0.0031 → -0.0034
Moderate / proportional	53.8% → 54.0%	+143.9 → +160.7	+7.0 → +13.4	-0.13 → -0.24	+0.0031 → +0.0029
Moderate / expansive	53.8% → 54.0%	+164.7 → +181.5	-19.7 → -13.5	+0.71 → +0.64	+0.0094 → +0.0092
Rapid / shares fixed	fixed	+575.8 → +575.8	+121.5 → +121.5	-1.09 → -1.09	+0.0025 → +0.0025
Rapid / compressive	51.3% → 52.0%	+172.9 → +233.0	+33.9 → +58.0	-1.54 → -1.64	-0.0048 → -0.0057
Rapid / proportional	51.3% → 52.0%	+211.4 → +271.8	-16.2 → +7.0	+0.11 → -0.17	+0.0073 → +0.0064
Rapid / expansive	51.3% → 52.0%	+252.1 → +313.4	-69.5 → -47.3	+1.72 → +1.59	+0.0198 → +0.0190
realization: Rapid / proportional / realization 0%	51.3% → 52.0%	-60.5 → +26.4	-16.2 → +7.0	+0.14 → -0.10	+0.0001 → -0.0000
realization: Rapid / proportional / realization 25%	51.3% → 52.0%	+6.7 → +86.9	-16.2 → +7.0	+0.13 → -0.12	+0.0019 → +0.0016
realization: Rapid / proportional / realization 50%	51.3% → 52.0%	+74.2 → +147.7	-16.2 → +7.0	+0.11 → -0.13	+0.0037 → +0.0032
realization: Rapid / proportional / realization 75%	51.3% → 52.0%	+141.8 → +209.4	-16.2 → +7.0	+0.13 → -0.13	+0.0055 → +0.0048
realization: Rapid / proportional	51.3% → 52.0%	+211.4 → +271.8	-16.2 → +7.0	+0.11 → -0.17	+0.0073 → +0.0064
capital_scope_excluding_retirement: Slow / proportional	55.0% → 55.0%	-3.9 → -3.9	-5.4 → -5.4	+0.05 → +0.05	+0.0006 → +0.0006
capital_scope_excluding_retirement: Moderate / proportional	53.8% → 54.0%	+98.3 → +118.0	+7.0 → +13.4	-0.11 → -0.21	+0.0026 → +0.0024
capital_scope_excluding_retirement: Rapid / proportional	51.3% → 52.0%	+105.9 → +176.7	-16.2 → +7.0	+0.15 → -0.12	+0.0061 → +0.0054

Findings that changed

Rapid aggregate labor income now grows 0.41% instead of falling 0.94%:

• Poverty under the tilt. Rapid / proportional poverty now falls 0.17 pp. On the July inputs it rose 0.11 pp. The tilt erodes most of growth's poverty gain (−1.09 pp with shares fixed) instead of reversing it.
• Realization. Revenue stays positive at every realized share: +$26B at 0%, where the July inputs gave −$60B and a breakeven at 22.5%.
• Retirement in the capital set. Excluding it cuts the Rapid estimate by about a third ($272B → $177B), not half.
• Tax-base choices vs scenario range. The two tax-base choices span $245B of the Rapid estimate, nearly the $265B between Slow and Rapid. On the July inputs they spanned more.
• Bridge. The bridged headline is +$344B against their +$246B (1.39×; was +$301B vs +$216B).
• Payroll. It agrees to within $5.1B in the nine as-forecast cells (Rapid: ours +58/+7/−47 against theirs +54/+5/−48 for compressive/proportional/expansive). That is agreement on the cap-and-spread channel. With shares fixed, where labor income grows with output, ours is 1.4× theirs (+$121B against +$84B in Rapid); the paper and memo now say so.
• Average-tax-rate cross-check, now like for like. Their Rapid band (21–32%) is federal revenue with their wedge over the $767B of gross factor income their shock adds: 21/26/32% by variant. On the same numerator, our cells run 30/34/39% of the market income ours adds, higher in every cell as the bridge is. The old text set our household-basis 30.7% against the band of their variants.
• Tilt vs growth. The intro said the tilt, not the growth rate, "drives" the results. Whether it outweighs growth now depends on the decomposition order. The finding now states what holds either way: the tilt takes back much of what growth delivers (47% of Rapid's revenue gain kept), and more of the poverty gain (16% kept).

The paper and memo carry a dated revision note, and every sentence these results falsified is rewritten. The /income-shift panel has a revision notice linking to the memo. "The great majority of workers" below the Rapid / expansive crossover ($139,212) still holds: in the build-p input microdata, 85–92% of workers earn below it for any wage growth to 2030 up to 40%.

How it is checked

• test_ai_scenarios.py
  • Both vintages of the published Table 1, to 0.005 pp.
  • A differential test against AI-Fiscal's committed cell_params for all 18 cells in both releases, to 1e-12.
  • Hypothesis property tests: the factor accounting identity θ0_K·g_K + θ0_L·g_L = g_Y, post-shock shares, monotonicity in the target share, fixed shares and λ symmetry.
• test_scenario_vintages.py
  • Same runtime, data and accounting; identical baseline.
  • Every row carries exactly the scenario the module builds today.
  • Untouched cells move by less than $0.1B; every reallocated Moderate and Rapid cell moves by more than $1B.
• test_paper_claims.py: each qualitative sentence in the paper is checked against the generated values ("erodes most", "nearly doubles", "about a third", "positive at every realized share", ...), plus the one paper value the site restates.
• test_memo_tables.py: every memo table and every quoted memo figure are checked against the run and the paper macros, and memoContent.js must be generated from the memo.
• test_generated_files_in_sync.py: the site payload, the state exposure deltas, and the paper's values and tables must regenerate exactly from the run. Both Budget Lab workbooks are pinned to AI-Fiscal's git blobs (b4cdf12c v1.0.1, 3090e94d tag v1.0.0). CI now installs openpyxl, so the regeneration check runs there.
• test_head_start_correction_outputs.py and the paper's September correction note now measure that correction on the July-inputs run (ai_scenarios_v1_0_0_inputs.json), so the revision cannot leak into it.
• Checkpoint keys now carry a digest of the scenario and shock settings (year, capital and labor sets, pass-through share), and the baseline is keyed the same way. Before, a resumed run would have reused the July Moderate and Rapid cells, because the revision changed no label.
• pytest analysis/tests: 259 passed on the 5.0.1 engine environment; 258 passed and 1 skipped (needs policyengine_us) in the CI environment. vitest: 41 passed. eslint --max-warnings=0: clean. The paper PDF builds with no undefined references; the eight overfull boxes are the same ones as on main.

Invariants

• Factor accounting: θ0_K·g_K + θ0_L·g_L = g_Y, and labor's post-shock share equals θ1_L, for all inputs (Hypothesis).
• Monotonicity: a lower target labor share moves growth from labor to capital at fixed g_Y. Fixed shares give g_K = g_L = g_Y. λ_compressive + λ_expansive = 2.
• Differential: g_Y, g_K, g_L and λ equal AI-Fiscal's R-code values for all 18 cells in both releases (1e-12).
• Vintage: same runtime ⇒ identical baseline. Only the reallocated Moderate and Rapid rows change inputs.
• Determinism: realization:1.0 equals grid Rapid / proportional exactly, though computed in separate processes.

Review

Three independent Opus reviewers (in session; Subfleet had no dispatchable lanes) covered replication and provenance, every claim and number, and what the PR missed. They found no blockers, and every number they recomputed matched. Their findings are applied in dc9b555. A fourth reviewer checked that delta: no blockers, and every recomputed figure correct (including the matched ATR, 21.05/26.17/32.11% theirs and 29.76/34.12/38.74% ours). Its three should-fix items, all about which comparison a sentence names, are applied in 7267996.

Run notes

One simulation of this pipeline peaks at about 80 GB of footprint, and a process does not release it all between units. The September run predates the fleet's memory caps. --max-new-units 1 now runs one unit per process, resuming from the input-keyed checkpoint, under heavy_run.sh with a 110 GB cap. The first process peaked near 81 GB and no later one passed 91 GB; the grid took about an hour.

Out of scope

• ai-inequality-slides and ai-inequality-populace predate the July report and quote none of its numbers. policyengine-app-v2 does not reference these routes.
• The mechanism sweep (shiftSweepData.json) is untouched by the revision and still predates the September Head Start correction, as the page says.

Files

• analysis/outputs/ai_scenarios.json: the rerun. ai_scenarios_v1_0_0_inputs.json is the 26 September run, byte for byte.
• analysis/outputs/yale_publishable_2030.xlsx: AI-Fiscal's v1.0.1 workbook, equal to run D of their results/erratum_v1.0.1. yale_publishable_2030_v1_0_0.xlsx is the July one.
• src/data/aiScenariosData.json: byte-identical to ai_scenarios_website_payload() of the run. It now carries each scenario's survey inputs, which the panel reads instead of restating them.
• src/data/stateExposureData.json: compute_state_exposure.build() on the run's checkpointed per-state baseline. Baselines are unchanged (Oklahoma differs by 3e-11 $B).
• Paper values, tables, figures and PDF regenerated. src/data/memoContent.js comes from bun run build:memo.

axiom: n/a: replication-calibration and output refresh, no policy encoding

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
ai-inequality	Ready	Preview	Oct 17, 2026 12:34am UTC

[MaxGhenis]

Hand-off from #51, which fixes the site's method copy. #51 leaves the paper's wording to this PR, since this PR regenerates the paper and its PDF. Three spots on this PR's head (7267996) carry wording #51 corrects on the site:

• paper/sections/methodology.tex:55 says the sweep distributes the freed labor income "to positive capital income in proportion to existing holdings". The sweep's shift (labor_capital_shift._apply_shift) scales every positive amount of the capital income flows by one common factor, so each person gains in proportion to the capital income they already receive, and losses don't change.
• paper/sections/methodology.tex:124-125 (:115-116 on main) uses the same "in proportion to existing holdings" for the calibrated scenarios' capital shock; check it against the scenarios' own allocation code before rewording.
• paper/sections/results.tex:166 (:164 on main) says "State-funded benefit responses" and that the state-benefit registry at the paper's model version (1.764.6) "covers state SSI supplements and child-care subsidies". At 1.764.6, household_state_benefits also lists other programs, such as New Jersey's property tax relief, California's CAPI, CARE and FERA, and Massachusetts's TAFDC. It leaves out Medicaid, CHIP and every other state's TANF.

Merge order: this PR rewrites the AIScenarios.jsx revenue paragraph and regenerates src/data/stateExposureData.json, and #51 changes both (in the data file, only its scope_note, which compute_state_exposure.py's SCOPE_NOTE writes). Whichever PR merges second should keep #51's copy.

Separately, at 1.764.6 household_benefits counts Massachusetts's TAFDC twice, once through tanf and once through household_state_benefits (a one-household run gives $19,072 for $9,536 of TAFDC). PolicyEngine/policyengine-us#9478 fixed it in policyengine-us 2.4.4, so the paper's net revenue totals carry Massachusetts's TAFDC changes twice until the study moves to 2.4.4 or later.

(Edited: the first version gave main's line numbers; the TAFDC paragraph came later.)