Net the 2026 section 68 taxable-income leg of the QBI and other non-itemized deductions

[MaxGhenis]

Fixes #9170.

Independent review: GPT-6.1 Sol approved (r1 approve with nits at 0624217ff6, follow-ups applied; r2 approve at adaa40787b; r3 approve at 50c302cee8).

What changes

From 2026, section 68(a) as rewritten by P.L. 119-21 sec. 70111(a) reduces itemized deductions by 2/37 of the lesser of "(1) such amount of itemized deductions, or (2) so much of the taxable income of the taxpayer for the taxable year (determined without regard to this section and increased by such amount of itemized deductions) as exceeds the dollar amount at which the 37 percent rate bracket under section 1 begins". Main took leg (2) as AGI minus exemptions. For an itemizer, taxable income increased by the itemized deductions is AGI less exemptions and every deduction section 63(b) allows besides itemized deductions, which section 63(d)(2) keeps out of "itemized deductions". Those are the QBI (199A), tips (224), overtime (225), senior (151(d)(5)(C)) and car-loan interest (163(h)(4)) deductions.

1. taxable_income_before_itemized_deductions (new) is AGI − exemptions − every deductions_if_itemizing entry except the itemized deductions. Wagering losses are a section 165(d) itemized deduction listed apart from the others, so they are added back. The 2026 branch of itemized_taxable_income_deductions_reduction and Idaho's recomputation of the limitation (id_itemized_deductions) use it.
2. taxable_income_less_qbid deducts itemized deductions before section 68 whenever the P.L. 119-21 section 68 applies. Sec. 70111(b)(1) amends 199A(e)(1) to read "taxable income shall be computed without regard to section 68 and without regard to any deduction allowable under this section", with the same effective date. This also removes the cycle reduction → QBI deduction → its taxable income → reduction.
3. tax_unit_itemizes, while the P.L. 119-21 section 68 applies, takes itemized deductions from the branch that itemizes, the one tax_liability_if_itemizing already uses. The limitation now depends on the QBI deduction, which depends on the itemization choice. In branch mode this adds no branch. In other years, and in the prior-law counterfactual, it reads the main simulation's value as before.

For a federal non-itemizer, the main simulation's itemized_taxable_income_deductions, which states such as Idaho and Colorado read, now nets the QBI deduction as calculated with the standard deduction. The QBI deduction "if itemizing" could differ by at most 20% of the gap between itemized and standard deductions, so the reduction could differ by at most 2/37 of that. Either is closer to the statute than main's AGI less exemptions.

The full legal argument, with verbatim extracts (P.L. 119-21 secs. 70103 and 70111, 26 U.S.C. 63(b) and (d), 199A(e)(1), Rev. Proc. 2025-32 sec. 4.01, P.L. 119-108 sec. 2) and the 2026 IRS drafts that corroborate it (Schedule A line 18, Form 1040 lines 13a-13b, Schedule 1-A line 44), is in #9170 (comment).

Worked example (statute arithmetic)

2026, single, AGI $700,000 including $100,000 of qualified REIT dividends, $50,000 of itemized deductions:

	main	this PR (= statute)
199A taxable income	652,702.70 (after section 68)	650,000 (without section 68)
QBI deduction	20,000	20,000
leg (2): taxable income without section 68, plus itemized deductions	700,000	680,000
reduction: 2/37 × min(50,000, leg (2) − 640,600)	2,702.70	2,129.73
taxable income	632,702.70	632,129.73

With $300,000 of the AGI in REIT dividends, leg (2) is 640,000, so there is no reduction; main reduces by $2,702.70.

Which deductions matter today

Under current-law parameters only the QBI deduction moves results. Leg (2) can exceed a bracket start only when AGI does, and the lowest start is $384,350 (separate). By then the other 63(b) deductions are fully phased out:

• tips: $25,000 less $100 per $1,000 of MAGI over $150,000 ($300,000 joint);
• overtime: $12,500 ($25,000 joint), same phase-out;
• tips and overtime are joint-only for married filers;
• car-loan interest: $10,000 less $200 per $1,000 over $100,000 ($200,000 joint);
• senior: $6,000 less 6 percent over $75,000 ($150,000 joint).

They count under reforms that widen those phase-outs, and the tests cover them through direct inputs.

Not changed here

• The qualified net disaster loss (63(b)(8), added by P.L. 119-108 sec. 2(c), effective for taxable years beginning after 2024) is not an itemized deduction either. PolicyEngine has no path for it into the 2/37 base: casualty_loss_deduction is off from 2018 (gov.irs.deductions.itemized.casualty.active). Noted on Personal casualty losses are non-deductible from 2018 in PolicyEngine, but §165(h)(5) still allows losses attributable to federally (and, from 2026, State) declared disasters #9920 for whoever models disaster losses.
• The truncated 2/37 rate is §68 reduction uses a truncated 2/37 rate and float32 rounding #9171.
• Wagering losses in leg (1) and the 165(d) 90% limit are Count gambling winnings in federal gross income and fix the wagering losses deduction #10070. The two PRs overlap only textually, in two test files (Count gambling winnings in federal gross income and fix the wagering losses deduction #10070 (comment)).
• taxable_income_less_qbid does not subtract personal exemptions. That predates this PR and matters only in reforms that restore exemptions; a follow-up session is proposed for it.

Invariants

For every input, with the P.L. 119-21 section 68 in force:

• 0 ≤ reduction ≤ 2/37 × itemized deductions, so allowed itemized deductions ≥ 35/37 of them.
• The reduction equals 2/37 × min(itemized, max(0, AGI − exemptions − 63(b) deductions − bracket start)), with the 2026 bracket starts restated from Rev. Proc. 2025-32 sec. 4.01. This is a differential test against an independent restatement, which also checks the parameter file.
• Each 63(b) deduction lowers leg (2) one for one. It never raises the reduction and never lowers it by more than 2/37 of itself. Wagering losses leave leg (2) unchanged.
• More AGI never lowers the reduction. More of any itemized deduction never lowers allowed itemized deductions.
• With the QBI deduction calculated from qualified REIT dividends, it equals min(20% × dividends, 20% × max(0, AGI − itemized deductions before section 68)). The reduction nets it out of leg (2), whether the itemization choice is an input or calculated.
• Every entry point gives the same values in branch and non-branch itemization modes (checked with a script for the reduction, tax_unit_itemizes, taxable_income_less_qbid, the QBI deduction, Idaho itemized deductions and household_net_income, 2025 and 2026).

Tests

• YAML from statute arithmetic and Rev. Proc. 2025-32:

  • itemized_taxable_income_deductions_reduction.yaml: single, joint, separate and head of household, plus a REIT case where the 20%-of-taxable-income cap binds;
  • taxable_income_before_itemized_deductions.yaml;
  • taxable_income_less_qbid.yaml: 2026, and the 2026 prior-law counterfactual;
  • tax_unit_itemizes.yaml: a case the corrected limitation flips to itemizing.

  Seven of the new cases fail on main; the head of household case (leg (1) binds) and the counterfactual pass there, as designed.

• Property and differential tests: four new tests (five cases) in tests/core/test_itemized_deduction_limitation_properties.py. All fail on main.

• Cost of the property file: 28.2 s and 2.31 GB peak RSS on main, against 32.8 s and 2.76 GB on this branch. These are back-to-back local macOS runs under a host load of about 260, not CI. The new 2026 tests take about 1.3 s of call time.

Microsimulation

A real 2026 microsimulation on the default dataset, comparing main at the merge base ad02ee1b2f with this PR at 0624217ff6. The runs went one at a time under the shared heavy lock. The later commits change only tests and the non-OBBB, non-branching path of tax_unit_itemizes, which a 2026 run with default settings does not take. Sums are weighted; the record counts are of tax units.

Output	Change	Records changed (weighted)
itemized_taxable_income_deductions_reduction	−$35.5M (base $10.63bn)	90 (151,427), all down
itemized_taxable_income_deductions	+$35.5M	90 (151,427)
qualified_business_income_deduction	−$0.6M	7 (4,666), all down
taxable_income	−$32.1M	61 (135,758): 57 down, 4 up
income_tax	−$10.3M	61 (135,758)
state_income_tax	−$392	5 (717)
household_net_income	+$10.3M	60 (135,758)
tax_unit_itemizes, total_itemized_taxable_income_deductions, household_benefits	0	0

• The reduction falls for top-bracket itemizers with a QBI deduction.
• The QBI deduction falls slightly where its 20%-of-taxable-income cap binds, because that taxable income now ignores section 68.
• That is also why taxable income rises for 4 units.

The comparison files are fixes/10136-impact/compare.2026.json on the hub host.

YAML suites at 50c302cee8: the IRS shard (tests/policy/baseline/gov/irs, per subfolder) and the Idaho, Colorado, Georgia, South Carolina, Utah and Louisiana folders passed, 4,364 cases with 0 failures, under the shared heavy lock. The 15 other state files that read the changed variables (96 cases) passed locally.

axiom: TheAxiomFoundation/rulespec-us#1697 queued. us:statutes/26/68#itemized_deduction_reduction_base is right given its input. But the 2026 composition never derives taxable_income_determined_without_section_68, nor 199A's taxable_income_computed_without_section_199A, so two committed fixtures contradict 199A(a)(2) arithmetic. #1697 has the verbatim law, a pasteable review_finding and companion tests from the same arithmetic as the YAML here.

🤖 Generated with Claude Code

[MaxGhenis]

Hand-off to the US + core hub.

• Head: 50c302cee8391f485e8ddc11a76e721708eae790. Ready for review; mergeable: MERGEABLE. CI started on the ready-for-review run.
• Independent review: GPT-6.1 Sol via Subfleet (--task review --tier hard).
  • r1 at 0624217ff6: approve with nits, none blocking. The two code nits were applied in adaa40787b.
  • r2 at adaa40787b: approve. Its optional test nits were applied in 50c302cee8.
  • r3 at 50c302cee8: approve.
  • Reviews: ~/PolicyEngine/_drafts/irc68-ti-leg-2026-10-11/review/sol_review_r{1,2,3}.md.
• Passed locally:
  • New and changed YAML: 38 cases.
  • Property file: 13 tests.
  • IRS shard plus the ID, CO, GA, SC, UT and LA folders under the heavy lock: 4,364 cases, 0 failures.
  • The 15 other state files that read the changed variables: 96 cases.
  • The 2026 microsimulation is in the body: income tax −$10.3M; 90 records with a lower reduction.
• Pending: CI only.
• Max's judgment: none needed. This is a bug fix to the statute, and the change in aggregates is stated in the body.
• Depends on: nothing.
• Overlaps: Count gambling winnings in federal gross income and fix the wagering losses deduction #10070 overlaps in two test files, textual only; whichever lands second resolves it (Count gambling winnings in federal gross income and fix the wagering losses deduction #10070 (comment)).
• Axiom: Derive the 2026 section 68(a)(2) and 199A(e)(1) taxable-income boundaries in the taxable-income composition (PE #10136 parity) TheAxiomFoundation/rulespec-us#1697 (pe-parity), for the Axiom eng hub.
• Follow-ups owed:
  • taxable_income_less_qbid omits personal exemptions, pre-existing; a follow-up session is proposed.
  • The truncated 2/37 rate is §68 reduction uses a truncated 2/37 rate and float32 rounding #9171.

[MaxGhenis]

US + core hub merge audit, us#10136 at 50c302cee8391f485e8ddc11a76e721708eae790 (squash): Net the 2026 section 68 taxable-income leg of the QBI and other non-itemized deductions.

• Review: independent GPT-6.1 Sol review APPROVE at this head (round 3; rounds 1 and 2 approved with nits, since applied).
• Gates checked live at merge:
  • gh pr checks exits 0;
  • the latest Pull request run at the head concluded success;
  • MERGEABLE;
  • not a draft;
  • no reviewer's latest review requests changes;
  • the semantic-overlap guard;
  • head pinned.
• Impact: a full 2026 microsimulation in the description: income tax -$10.3M, from 90 records with a smaller section 68 reduction. A bug fix to the statute.
• Axiom: Derive the 2026 section 68(a)(2) and 199A(e)(1) taxable-income boundaries in the taxable-income composition (PE #10136 parity) TheAxiomFoundation/rulespec-us#1697 queued. us:statutes/26/68#itemized_deduction_reduction_base is right given its input. But the 2026 composition never derives taxable_income_determined_without_section_68, nor 199A's taxable_income_computed_without_section_199A, so two committed fixtures contradict 199A(a)(2) arithmetic. Add Washington state to modelled_policies.yaml #1697 has the verbatim law, a pasteable review_finding and companion