[MaxGhenis]

What this changes

We change copy only. After we drop their description strings, the four data files match main, so no number on the site changes.

• The shift. On main, we describe the shift as moving the freed labor income to capital "in proportion to existing holdings". labor_capital_shift._apply_shift scales every positive amount of the capital income flows in constants.CAPITAL_INCOME_VARS (long- and short-term capital gains, taxable interest, qualified and non-qualified dividends, rental income) by one common factor. uk_shift_sweep._apply_shift does the same for the UK's flows: savings interest, ISA interest, dividends, property income and capital gains. Each person then gains in proportion to the capital income they already receive, and losses stay as they are. We rewrote the US and UK sweep descriptions, the labor-shift description, their data files, two chart captions and the "The experiment" cards on the income-shift and policy-analysis pages to say that. Both pages serve the US and the UK, so on their cards we describe the common factor without listing one country's flows.

• The state measure. On main, we call the subtracted benefits "state-funded benefits", and a reader would take that to mean everything a state pays for. We subtract only household_state_benefits, the programs PolicyEngine counts as state-paid. At policyengine-us 1.764.6, the version behind the published state figures, PolicyEngine counts state SSI supplements, child care subsidies and other programs, such as New Jersey's property tax relief, as state-paid. It leaves Medicaid, CHIP and TANF off that list, apart from Massachusetts's TAFDC.

  New York's and Maryland's figures include New York City's income tax and refundable credits, Maryland's county income taxes and Montgomery County's earned income credit (md_montgomery_eitc). California's include San Francisco's Working Families Tax Credit (ca_sf_wftc, one of California's refundable credits) and three Los Angeles County programs in household_state_benefits: ca_la_ez_save, ca_la_infant_supplement and ca_la_expectant_parent_payment. We checked every refundable credit list and the household benefit lists against the variables under gov/local and found no others. We rewrote the state exposure scope note and the state ranking caption to say all this, and dropped "state-funded benefits" from the scenarios page's revenue note. We also cut the caption's claim that the chart shows only states with an income tax: in dollars, it ranks every state in state_deltas.

• The revenue label. On main, we label the sweep's revenue option "Net federal revenue change", but the chart plots revenue_change_b, an alias of total_rev_change_b, which nets federal, state and local taxes against transfers. We relabeled it "Net government revenue change", the UK sweep's label. We also rewrote its description and the scenarios page's revenue note to say we count local taxes in the total and leave out health coverage. At 1.764.6, household_health_benefits returns zero unless a run switches on gov.simulation.include_health_benefits_in_net_income, and we leave that parameter at its default, off.

• The federal chart. buildFederalRow adds up its own bands rather than taking the total less the states. At 1.764.6, employer_payroll_tax includes employers' state unemployment insurance and other state and local payroll taxes, so the federal payroll band holds them, and the benefits band includes TANF. We count three groups of taxes only in the total: workers' state payroll taxes such as California's SDI, the Additional Medicare Tax, and local taxes on income and occupations. PolicyEngine adds up those local taxes in local_income_tax_before_refundable_credits and local_occupational_tax. It counts New York City's and Maryland's local income taxes in state_income_tax_before_refundable_credits, so they land in those states' charts. On the published sweep at a 10% shift, the total less the states comes to -$138.2B, against the federal chart's -$131.3B. We added a note under the US federal chart to say so.

• Analysis notes. On main, in california_shift.py and california_summary.py we call "all governments less the states" the published site's federal line. We now call it the split california_summary applies to the published sweep, and in the module docstring we explain how the site's federal chart differs. We replaced "state-funded benefits" with household_state_benefits in compute_transfer_detail.py and compute_state_exposure.py.

• Voice. We gave the method copy and the data descriptions actors and shorter sentences. In them, we now move the income, run the current-law simulation and name what we leave out, and we write the stat card's and the state ranking's measures as equations. We split the federal chart note's 47-word sentence in two and the 41-word list of local items by state, and named the three Los Angeles County programs. We also cut "Reconciles to the household net-income identity" from the revenue description.

Some of the strings we fixed don't render on the site today: ShiftSweep doesn't read the sweep's metadata.description, no page renders LaborShift.jsx, and no page imports stateExposureData.json. We fixed them too, since the published JSON carries them and every future run rewrites them.

What this leaves alone

• The paper. We use the same wording in three places in the paper on main: paper/sections/methodology.tex:55 and :115-116 ("in proportion to existing holdings") and paper/sections/results.tex:164 ("State-funded benefit responses"). On Match the Budget Lab's 9 Oct 2026 revision: rerun scenarios, regenerate site, memo and paper #45's head (7267996), those lines sit at :55, :124-125 and :166. In Match the Budget Lab's 9 Oct 2026 revision: rerun scenarios, regenerate site, memo and paper #45 we regenerate the paper and its PDF, so we'll make the fixes there rather than rebuild the PDF by hand here. We left a comment on Match the Budget Lab's 9 Oct 2026 revision: rerun scenarios, regenerate site, memo and paper #45 to hand them off.
• Merge order with Match the Budget Lab's 9 Oct 2026 revision: rerun scenarios, regenerate site, memo and paper #45. In Match the Budget Lab's 9 Oct 2026 revision: rerun scenarios, regenerate site, memo and paper #45 we rewrite the same AIScenarios.jsx paragraph and regenerate stateExposureData.json from SCOPE_NOTE. Whichever PR we merge second, we'll keep this PR's copy when we resolve the conflict.
• The numbers. To put the uncharted taxes in a band, we have to store them separately in the sweep, so we leave that fix to the study's next refresh. For the refresh, we also need policyengine-us 2.4.4 or later: earlier releases count Massachusetts's TAFDC twice in household_benefits, once through tanf and once through household_state_benefits. Those include 1.764.6 and 2.2.1, the version the policyengine 6.16.0 bundle pins. In a one-household run, household_benefits returns $19,072 for $9,536 of TAFDC. PolicyEngine fixed the double count in 2.4.4 (Count MA TAFDC once in household benefits policyengine-us#9478), and california_shift.py already counts TAFDC once.

Checks

• bun run lint -- --max-warnings=0: clean.
• bun run test:run: 10 files, 36 tests passed. ShiftSweep.test.jsx now checks the revenue label, the health exclusion, the federal chart note and the state caption's list of local items.
• pytest analysis/tests -q on Python 3.11 with numpy, pandas, pytest and hypothesis, as CI installs: 188 passed, 13 skipped.
• CI at a2de7ff: lint-and-test and the Vercel preview build pass.
• Data files: apart from the description strings, all four match main, and SCOPE_NOTE matches the scope note in stateExposureData.json.
• Prettier: a2de7ff rewraps three lines of ShiftSweep.jsx, so it now passes, as IncomeShift.jsx and PolicyAnalysis.jsx do. AIScenarios.jsx keeps main's three Prettier differences, which sit outside this PR's lines.
• Remove retired dataset names from the study site #50: its head (0c6ee17) changes model_url in shiftSweepData.json, on the line after the description this PR rewrites, and the two edits conflict. When Remove retired dataset names from the study site #50 merges, we'll rebase onto main, keep Remove retired dataset names from the study site #50's model_url and this PR's description, and rerun these checks along with Remove retired dataset names from the study site #50's retired-name test.
• Independent review: the round-1 reviewer asked for changes at 9d91554, with one blocking finding on local taxes. The round-2 reviewer approved 61a3f4a with nits, which we applied in df33f2c, and the delta reviewers approved df33f2c and 12dda9d, which adds the remaining local items. In 8daa5c2 we added "credits" to that list's lead-in, as the 12dda9d review suggested, and its delta review found nothing. The round-4 copy review, which covers this PR and Count health coverage and TANF in the income-shift sweep (d2671) #53, has yet to report on 44d1c87 and a2de7ff.

We stacked #53 on this branch, and we'll keep this PR in draft until #50 merges.

🤖 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	Nov 10, 2026 7:42pm UTC