Cap the regular tax at the tax on all taxable income (26 U.S.C. 1(h)(1)) and use it on Form 6251 line 10

[MaxGhenis]

Follow-ups 1, 2 and 5 from #9792.

Summary

• Regular tax capped at the tax on all taxable income. 26 U.S.C. 1(h)(1) says the tax "shall not exceed" the sum of the subparagraph (A) to (F) amounts. Schedule D Tax Worksheet line 47 takes "the smaller of line 45 or line 46", where line 46 is the tax on all of line 1 at the regular rates. capital_gains_tax now stops at tax_on_taxable_income_at_main_rates (new, line 46) less income_tax_main_rates. The Biden and Harris capital gains reforms apply the same limit, and additional_tax_bracket figures line 46 on its own schedule.
• Form 6251 line 10 is the regular tax the model charges. regular_tax_before_credits was a second Schedule D Tax Worksheet. It returned the smaller of its own total, which included the gains, and income_tax_main_rates, which leaves them out. alternative_minimum_tax then added capital_gains_tax on top. regular_tax_before_credits is now income_tax_main_rates + capital_gains_tax, the same regular tax income_tax_before_credits adds, and line 10 uses it alone. Its meaning changes: it now includes the capital gains tax (see "Outside readers" below).
• Rev. Proc. 2025-32 citations. The 2026 items are section 4; section 3 changes only two 2025 items. Fixed in amt_tax_including_cg.yaml (3.10 → 4.10), the Maine aged-or-blind parameter (3.14(3) → 4.14(3)) and a test comment (".14(2)" → 4.14(2)). The repo's eleven other citations of the Rev. Proc. by section or name are right: 4.01 three times, 4.14(2) four times, 4.31 twice, and two without a section. Of the 22 parameter links to its PDF, three pointed into the wrong section and now point at the pages with their 2026 amounts. The capital gains rate thresholds moved from page 12 (section 4.02) to page 13 (section 4.03). The child tax credit amount and refundable maximum moved from page 13 to page 14 (section 4.05). The other links land on the page with the amounts or on the page where their section starts.

axiom: TheAxiomFoundation/rulespec-us#1501 queued

Law and worksheet lines

26 U.S.C. 1(h)(1): "If a taxpayer has a net capital gain for any taxable year, the tax imposed by this section for such taxable year shall not exceed the sum of— (A) a tax computed at the rates and in the same manner as if this subsection had not been enacted on the greater of— (i) taxable income reduced by the net capital gain; or (ii) the lesser of— (I) the amount of taxable income taxed at a rate below 25 percent; or (II) taxable income reduced by the adjusted net capital gain; (B) 0 percent ... (F) 28 percent of the amount of taxable income in excess of the sum of the amounts on which tax is determined under the preceding subparagraphs".

2025 Schedule D Tax Worksheet, lines 44 to 47:

• line 44: "Figure the tax on the amount on line 21" (the subparagraph (A) amount);
• line 45: "Add lines 31, 34, 40, 43, and 44";
• line 46: "Figure the tax on the amount on line 1";
• line 47: "Tax on all taxable income (including capital gains and qualified dividends). Enter the smaller of line 45 or line 46".

The Qualified Dividends and Capital Gain Tax Worksheet does the same on lines 23 to 25. 2025 Form 6251 line 10: "Add Form 1040 or 1040-SR, line 16 (minus any tax from Form 4972), and Schedule 2 (Form 1040), line 1z. Subtract from the result Schedule 3 (Form 1040), line 1 ...". Form 1040 line 16 is the worksheet's line 47.

Worksheet numbering. I compared the worksheet text for every year from 2017 to 2025:

• The 2019 to 2025 worksheets have the same lines and instructions. Only the dollar amounts and the Form 1040 line numbers they cite differ.
• The 2018 worksheet has lines 18a, 18b, 18c and 19 where 2019 has lines 18 to 21. From there on, 2018 line N is 2025 line N + 2: 2018 line 26 is 2025 line 28, and 2018 line 45 is 2025 line 47.
• The 2017 worksheet has the same numbering from line 19 on. Its line 19 is the larger of lines 17 and 18 because, before 2018, the 0 percent amount was the top of the 15 percent bracket.

The old regular_tax_before_credits used the 2017/2018 numbering.

Where line 45 exceeds line 46. It happens only where gain taxed at 15 percent sits in the 12 percent bracket. The 0 percent rate amount ends below the top of that bracket by $100 to $250 in 2018 to 2025 (up to $3.75 single and $7.50 joint). Since P.L. 119-21 it is $950, $1,900 and $1,250 in 2026 (Rev. Proc. 2025-32 sections 4.01 and 4.03), worth up to $28.50 single and separate, $57 joint and $37.50 head of household. The 25 and 28 percent amounts are always above the 24 percent bracket top, where the regular rate is at least 32 percent, so line 45 never exceeds line 46 because of them.

Traced cause of follow-up 1

#9792's run found regular_tax_before_credits below income_tax_main_rates for 18 tax units in 2025 ($328M weighted, up to $19,123). The run used #9780's head, before #9781. All 18 units elect under Form 4952 line 4g (investment_income_elected_form_4952, from the dataset). Two errors stacked:

1. dwks10 ignored the election, while net_capital_gain applied it. dwks10 read investment_income_form_4952, an input no dataset fills. So worksheet line 10 was too large and line 21 (dwks19) too small. For unit 1101624, line 10 was 222,465 against a net capital gain of 80,593. Apply the Form 4952 election in Schedule D Tax Worksheet lines 3-10 #9781 (2137342) fixed this. On main, dwks10 equals net_capital_gain for every tax unit, to float32 rounding (at most $0.25). This PR ran main at ea79c39 for 2025 and 2026, and at 47ab53e for 2025. In those runs, no tax unit has regular_tax_before_credits below income_tax_main_rates by more than half a cent. The hub's run on 3f15391 found the same. The duplicate worksheet still fell below income_tax_main_rates by float32 rounding, at most $0.002. That happened for 789 tax units in 2025 and 701 in 2026, and for 790 in 2025 on 47ab53e.
2. The duplicate worksheet took line 28 as min_(line 21, line 22). Every worksheet adds them (2025 line 28 = lines 21 + 22; 2017 and 2018 line 26 = lines 19 + 20). Because min_ took line 22 instead of adding line 21, line 29 was too large, so gain that belonged at 20 percent was taxed at 15. For unit 1101624 that was all of line 23: $11,123 of the $19,123 gap (5% of 222,465). The other $8,000 came from cause 1.

Form 6251 line 10 used min(duplicate worksheet tax, income_tax_main_rates) + capital_gains_tax, so any shortfall in the duplicate went straight to line 10. Error 2 never showed on main, because the duplicate's total, gains included, stayed above income_tax_main_rates. The duplicate had other latent errors too:

• its line 46 was income_tax_main_rates, not the tax on line 1;
• it took the 25 and 28 percent rates from the AMT parameters;
• it ignored the skip instructions.

The fix removes the duplicate. Line 10 is now built from the same variables as income_tax_before_credits, so a difference between two routes can no longer change it. The worksheet itself is now transcribed in the tests (see below), where a disagreement fails CI instead of moving the AMT. None of the 18 units owed AMT in either run: line 9 stayed well below line 10.

Invariants (property tests, test_regular_tax_before_credits.py)

These hold for every household, filing status and year. The tests check them in three ways:

• on a grid of 280 households and on boundary households, in 2025 and 2026;
• on derandomized Hypothesis draws, 10 examples per test: batches of up to 25 households in 2025 and 2026, which are also compared with the transcription;
• on batches of up to 15 households in 2018, 2021, 2024, 2025, 2026, 2027, 2030 and 2035.

The invariants:

1. 0 <= regular_tax_before_credits <= tax_on_taxable_income_at_main_rates ("shall not exceed").
2. The preferential computation never exceeds the ordinary one: 0 <= capital_gains_tax <= line 46 - income_tax_main_rates, so income_tax_main_rates <= regular_tax_before_credits.
3. Identities:
   • regular_tax_before_credits = income_tax_main_rates + capital_gains_tax;
   • income_tax_before_credits = regular_tax_before_credits + alternative_minimum_tax;
   • the AMT is Form 6251 line 9 less line 10, if more than zero.
4. Without net capital gain, the regular tax is line 46, bit for bit.
5. Worksheet lines:
   • dwks10 is the net capital gain and dwks13 the adjusted net capital gain;
   • line 18 <= line 21 <= line 14;
   • income_tax_main_rates taxes line 21.
6. More wages, long-term gain or qualified dividends never lower the regular tax.
7. A household's results do not depend on the rest of its batch or their order.

The tests also cover every year from 2018 to 2035 and every filing status. Taxable income is set at the top of the 12 percent bracket, with dividends from the 0 percent amount up. In each case the limit binds and the gains tax is 12 percent of the dividends. The band is $100 to $250 wide in 2018 to 2025 and $950 to $1,900 in 2026.

Differential. Lines 1 to 47 of the 2025 Schedule D Tax Worksheet are transcribed line by line. The transcription includes:

• every skip instruction;
• Form 4952 lines 4d, 4e and 4g;
• capital gain distributions on Schedule D line 13, and short-term losses;
• Schedule D lines 18 and 19.

The regular rates come from the 2025 Tax Computation Worksheet's rows, also below $100,000, where the form uses the Tax Table, which rounds within $50 rows. The model does the same. 2026 runs on Rev. Proc. 2025-32's tables (sections 4.01 and 4.03), and a test checks the rows against the printed subtraction amounts and bases. The model matches the transcription's lines 10, 13, 14, 21, 44, 46 and 47 for:

• a grid of 280 households;
• boundary households with ordinary income at the 0 percent amount and dividends reaching the top of the 12 percent bracket, ± $1;
• Hypothesis draws, for both years.

The existing differential tests changed in two ways:

• test_form_6251_part_iii.py now compares the regular tax with Schedule D Tax Worksheet line 47.
• test_section_911_tax_stacking.py and test_form_4952_section_911_interaction.py compare it with Foreign Earned Income Tax Worksheet line 6, figured from line 25 of the Qualified Dividends and Capital Gain Tax Worksheet.

Both match to single-precision tolerance. Before, they matched line 45, or line 6 figured from line 23. A separate check allowed the capped figure to be up to $7.50 lower (in the Form 4952 test, up to the rate overlap).

A counterexample the properties found, now fixed on main by #9788. Gross income used to add non_sch_d_capital_gains outside the Schedule D loss limit, while net_capital_gain (since #9781) nets the same distributions on Schedule D line 13. With a Schedule D net loss, more long-term gain then raised net capital gain but not taxable income. A 2018 single filer with $50,000 of distributions, a $40,273 short-term loss and $1 of long-term gain paid $0.12 less. Before #9788 merged, property 6 set such households aside. #9788 nets the distributions inside the capital loss limit, so after merging main the property holds for every household. The minimized case is now an explicit Hypothesis example. It failed before #9788 and passes now.

Merged main

The branch merges main twice, with merge commits: at ea79c39 (3ec4485) and at 47ab53e (ede9aad).

• Give the additional_tax_bracket reform's unset bracket zero tax instead of NaN #9831. amounts_in_brackets is the shared bracket walk. The additional_tax_bracket reform's line 46 override calls tax_at_main_rates, so it uses that walk too. Give the additional_tax_bracket reform's unset bracket zero tax instead of NaN #9831's reform test listed regular_tax_before_credits among the variables the reform replaces. The reform now replaces tax_on_taxable_income_at_main_rates instead, so the test's default-system leak check covers the three replaced variables, income_tax_main_rates, tax_on_taxable_income_at_main_rates and taxable_income_taxed_below_25_percent. Its finite and monotone checks cover the reform's three taxes: income_tax_main_rates, tax_on_taxable_income_at_main_rates and regular_tax_before_credits.
• Net capital gain distributions inside the capital loss limit in AGI #9788. It removes the property-test exclusion above.
• Net losses against 28 percent rate gain and unrecaptured section 1250 gain (26 U.S.C. 1(h)(4), 1(h)(6)) #9841 nets losses against the 28 percent rate gain and unrecaptured section 1250 gain. Only test_section_911_tax_stacking.py changed on both sides, in different functions, and git merged it without conflict. Net losses against 28 percent rate gain and unrecaptured section 1250 gain (26 U.S.C. 1(h)(4), 1(h)(6)) #9841 passes Schedule D lines 18 and 19 entered already netted through unchanged, so the transcription, which takes them as entered, still applies. This branch's capital_gains_tax and reforms read the section 911 amounts that Net losses against 28 percent rate gain and unrecaptured section 1250 gain (26 U.S.C. 1(h)(4), 1(h)(6)) #9841 updated.
• Keep dependents' above-the-line deductions off the filer's AGI #9801, Net Michigan business and rental losses in household resources (MI-1040CR lines 16, 17 and 30) #9790, Fix Arizona property tax credit Schedule 2 thresholds #9779, List casetext.com in the known-dead reference URL guard #9768, Limit the CTC by actual tax liability and make formula branches calculate under their overrides #9741 and Exclude input dependents from tax unit head and spouse inference #9630 touch none of this branch's files.

Tests, before and after

The new cases are in alternative_minimum_tax/regular_tax_before_credits.yaml and before_credits/tax_on_taxable_income_at_main_rates.yaml. The copies run on main and before #9781 leave out tax_on_taxable_income_at_main_rates, which does not exist there.

Cases	Main (ea79c39)	Before #9781 (94c4453)	This branch
4 dataset tax units (1101624, 1004930, 1156248, 1032460)	2 fail (1101624, 1156248) on regular_tax_before_credits, from the meaning change. 1004930 and 1032460 have no capital gains tax and pass.	4 fail, on dwks10 or net_capital_gain (the Form 4952 election defects #9781 fixed)	pass
7 cases where line 45 exceeds line 46 (2025 single, joint and Form 2555; 2026 single, joint, separate and head of household)	fail on capital_gains_tax by $3.75, $7.50, $3.75, $28.50, $57, $28.50 and $37.50	same	pass
3 cases where line 45 stays below line 46	fail on regular_tax_before_credits (meaning change only)	same	pass

Main: 12 failed, 2 passed. Before #9781: 14 failed. The failure lines on main are identical at ea79c39 and at the original base a2f0fd4. A third run, at 47ab53e, got 3 seconds of CPU in 9.5 minutes on the loaded host and was stopped.

Run on the branch after the second merge (ede9aad), one process at a time:

• capital_gains and before_credits: 91 passed, including Net losses against 28 percent rate gain and unrecaptured section 1250 gain (26 U.S.C. 1(h)(4), 1(h)(6)) #9841's 30 new cases;
• alternative_minimum_tax and foreign_earned_income_exclusion: 145 passed;
• the contrib reforms that override these lines (additional_tax_bracket, harris/capital_gains, biden/budget_2025, crfb/agi_surtax.yaml, congress/watca.yaml) and Net capital gain distributions inside the capital loss limit in AGI #9788's schedule_d_capital_gain_distributions.yaml: 88 passed;
• test_regular_tax_before_credits.py, test_form_6251_part_iii.py, test_section_911_tax_stacking.py, test_form_4952_section_911_interaction.py and test_negative_main_rates_inverted_brackets.py: 37 passed;
• Net losses against 28 percent rate gain and unrecaptured section 1250 gain (26 U.S.C. 1(h)(4), 1(h)(6)) #9841's test_capital_gain_rate_group_netting.py and Net capital gain distributions inside the capital loss limit in AGI #9788's test_capital_gain_distributions_loss_limit.py: 12 passed;
• Give the additional_tax_bracket reform's unset bracket zero tax instead of NaN #9831's test_additional_tax_bracket_unset_bracket.py: 9 passed.

The same suites passed after the first merge, and the YAML folders passed before either merge (275 tests).

Impact (default dataset)

Four full microsimulations were run: main at ea79c39 and this branch's change on it (d5f4e88), for 2025 and 2026. Each used populace_us_2024 (populace-us-2024-spm-20260909) and policyengine-core 3.32.15, with 79,729 tax units.

	2025	2026
Tax units where the line 46 limit binds (capital_gains_tax falls by more than half a cent)	2 (combined weight 0.7)	34 (44,140 weighted; 25 joint, 7 single, 2 head of household)
Largest cut for one tax unit	$1.23	$47.29
income_tax, weighted change	−$0.06	−$616,041
household_net_income, weighted change	+$0.06	+$616,120
alternative_minimum_tax	unchanged in every tax unit	unchanged in every tax unit
AMT payers (both runs)	110 (330,771 weighted)	170 (573,820 weighted)
state_income_tax	unchanged	unchanged

Every tax unit where the limit binds has taxable income near the top of the 12 percent bracket. In 2026, joint units are at $99,095 to $100,906, around the $100,800 top and above the $98,900 zero rate amount. Single units are at $49,576 to $50,480, and heads of household at $66,904 to $67,417. In 2025, the two units are at $96,757 (joint) and $48,439 (single). In 2025 one more tax unit's capital_gains_tax falls, and in 2026 five more, each by under $0.001 (float32). In both years, income_tax is bit-for-bit unchanged in every tax unit whose capital_gains_tax is unchanged. No tax unit's capital_gains_tax rose.

regular_tax_before_credits changes for 13,275 tax units in 2025 and 13,372 in 2026, by +$279.6 billion and +$294.0 billion weighted. That is the meaning change: it now includes the capital gains tax. Apart from float32 rounding, Form 6251 line 10 changes only where the limit binds, and by the same amount as the regular tax.

These invariants hold on every tax unit of the branch, in both years, to 5 cents plus 5e-7 of taxable income (the no-gain identity bit for bit):

• 0 <= regular_tax_before_credits <= line 46;
• regular_tax_before_credits = income_tax_main_rates + capital_gains_tax;
• 0 <= capital_gains_tax <= line 46 - income_tax_main_rates;
• income_tax_before_credits = regular_tax_before_credits + alternative_minimum_tax;
• dwks10 = net_capital_gain;
• without net capital gain, the regular tax is line 46;
• AMT = max(0, line 9 − line 10), for every tax unit (the dataset has no foreign earned income exclusion, foreign tax credit or Form 4972 amounts).

2025 again after the second merge. Main at 47ab53e against the branch at ede9aad gives the same result. The limit binds for the same 2 tax units (combined weight 0.7), income_tax falls by $0.06 weighted, the AMT is unchanged in every tax unit, and the invariants hold. regular_tax_before_credits changes for 13,278 tax units (+$279.6 billion). The matching 2026 run of the branch hit this build's 10-minute cap for a single run, after it had computed income_tax. Retries could not get the shared microsimulation lock, and the host's load average later passed 100. So the 2026 figures above are from ea79c39. #9841 passes the dataset's Schedule D lines 18 and 19 through unchanged.

Outside readers of regular_tax_before_credits

GitHub code search across the PolicyEngine and TheAxiomFoundation organizations (2026-10-06) finds one program outside this repo that reads the variable from policyengine-us. PolicyEngine/salt-amt-calculator's API (api/salt_amt_api/simulation/calculation.py) reports regular_tax_before_credits for 2026 as its "regular tax". It compares that with amt_base_tax (gap = max(regular_tax - amt, 0), amt_binds = amt > regular_tax). Its Modal image installs policyengine-us>=1.0.0, unpinned (api/modal_app.py), so its next image build picks up this change. api/uv.lock, which pins 1.485.0, is not used there. With this change, its regular tax for a household with long-term gains or qualified dividends includes the capital gains tax, as Form 6251 line 10 does. Its gap (max(regular_tax - amt, 0)) can grow by up to capital_gains_tax, and amt_binds can turn false. Its reported regular tax then matches Form 6251 line 10. The other matches are:

• documentation and audit notes (code-snippets, obbba-paper, policybench);
• Axiom's own regular_tax_before_credits in rulespec-us and the tools that read it. That variable is the subject of rulespec-us#1501.

No partner contract test under tests/policy/baseline/partners is touched.

Not in this PR

• Follow-up 3 (Schedule D lines 18 and 19 without the netting of 1(h)(4)(B) and 1(h)(6)) is Net losses against 28 percent rate gain and unrecaptured section 1250 gain (26 U.S.C. 1(h)(4), 1(h)(6)) #9841.
• Follow-up 4 (additional_tax_bracket NaN with default 2026 parameters) was fixed by Give the additional_tax_bracket reform's unset bracket zero tax instead of NaN #9831, merged here.

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the US + core hub on its merge gates. Audit:

• Head: ede9aad649.
• Independent review: GPT-6.1 Sol on Codex, ~/reviews/us-hub/reviews-gen/us-9852/review-r1.md. APPROVE at this head, with no P1 or P2 findings and no methodology choice.
  • The reviewer checked lines 45–47 of the Schedule D Tax Worksheet, the §911(f) stacking and Form 6251 line 10.
  • It recomputed the 2026 single case at the line 46 limit ($5,800) and a 2025 Form 2555 case.
  • It audited every reader of regular_tax_before_credits and found no double count in any state or reform.
  • Its one P3, the body wording on the calculator gap, is applied.
• What it fixes (follow-ups 1, 2 and 5 from Tax 28% rate and unrecaptured 1250 gain at regular rates below the 24% bracket top, and use Schedule D Tax Worksheet line 21 on Form 6251 line 27 #9792):
  • The regular tax is capped at the tax on all taxable income (line 47, §1(h)(1)).
  • Form 6251 line 10 uses that capped regular tax.
  • The duplicate worksheet with the line 28 min_ error is removed.
  • Rev. Proc. 2025-32 citations and anchors are corrected.
• CI: every check passed at this head.
• Impact: default dataset, 2026: the limit binds for 34 tax units (44,140 weighted), income tax falls $616,041, the AMT is unchanged, and the largest cut is $47.29. 2025: 2 units, −$0.06.
• Downstream: regular_tax_before_credits now includes the capital gains tax, as Form 6251 line 10 does. PolicyEngine/salt-amt-calculator's API reports it as "regular tax", so its next image build shows the corrected figure.
• Order: 01265359's test-only PR is stacked on this one and lands next.