[MaxGhenis]

The contributed repeal of state dependent exemptions (gov.contrib.repeal_state_dependent_exemptions) gave a Hawaii or Virginia filer whom another taxpayer can claim as a dependent the exemption that the baseline denies. An unclaimed but claimable single filer aged 22 got $1,144 in Hawaii and $930 in Virginia under the repeal, against $0 in the baseline. The repeal should remove dependents and nothing else, so it must never raise such a filer's exemption.

The US + core hub found this while reviewing #10044, which fixes the same defect for Rhode Island and Wisconsin. Both bad overrides are already on main and outside #10044.

The law

• Hawaii, HRS 235-54(a) (HRS ch. 235, PDF page 45): exemptions are the number the individual can claim under the Internal Revenue Code, plus one for a taxpayer or spouse aged 65 or over, times $1,144. "In the case of an individual with respect to whom an exemption under this section is allowable to another taxpayer ..., the personal exemption amount applicable to such individual under this subsection ... shall be zero." The aged exemption is part of the same subsection, so it is zero too.
• Virginia, Va. Code 58.1-322.03(2)(a) (section): "A deduction in the amount of $930 for each personal exemption allowable to the taxpayer for federal income tax purposes." Under IRC 151(d)(2), a filer whom another taxpayer can claim has no federal personal exemption. Subdivision (2)(b)'s $800 for each blind or aged taxpayer has no such bar.

The baseline already applies these rules (hi_regular_exemptions, va_personal_exemption_person). The repeal's overrides counted every head and spouse instead (head_spouse_count, is_tax_unit_head_or_spouse).

Change

In repeal_state_dependent_exemptions.py:

• Hawaii: the filers' own count is head_spouse_count_not_dependent_elsewhere, and the aged add-on counts aged & head_or_spouse & ~claimable_as_dependent_on_another_return. That is the baseline's own and aged gate. Only the dependent count is dropped.
• Virginia: each filer gets $930 only if not claimable (filer & ~claimable_as_dependent_on_another_return), the baseline's filer gate without its dependent term. va_aged_blind_exemption is not overridden, so it is unchanged.

This uses the same helper as #10044's RI and WI fix and the existing HI dependent-exemption reform.

Non-claimable returns. Hawaii's count moves from head_spouse_count (2 if filing_status is JOINT, else 1) to the number of heads and spouses in the unit who cannot be claimed. With no claimable filer, the two agree whenever filing status is computed, because the filing_status formula returns JOINT exactly when the unit has a spouse (tax_unit_married). They can differ only when a filing_status input disagrees with the unit's members. In that case the new count follows the baseline, which counts members (exemptions_count returns tax_unit_size for a unit with no claimable filer).

Before and after (2025)

Case	Output	Baseline	Repeal on main	Repeal, this PR
HI single, unclaimed but claimable, 22	hi_regular_exemptions	0	1,144	0
HI joint, one claimable spouse, 2 children	hi_regular_exemptions	1,144	2,288	1,144
HI joint, claimable spouse within the filing exception, 2 children	hi_regular_exemptions	3,432	2,288	1,144
HI joint, both spouses claimable	hi_regular_exemptions	0	2,288	0
HI single, claimable, 66	hi_regular_exemptions	0	2,288	0
HI joint 67/70, the 67-year-old claimable	hi_regular_exemptions	2,288	4,576	2,288
HI joint 70/68 with a child, no one claimable (control)	hi_regular_exemptions	5,720	4,576	4,576
VA single, unclaimed but claimable, 22	va_personal_exemption	0	930	0
VA joint, one claimable spouse, 2 children	va_personal_exemption	930	1,860	930
VA joint, claimable spouse within the filing exception, 2 children	va_personal_exemption	2,790	1,860	930
VA joint, both spouses claimable	va_personal_exemption	0	1,860	0
VA single, claimable, 66	va_total_exemptions	800	1,730	800
VA joint with 2 children, no one claimable (control)	va_personal_exemption	3,720	1,860	1,860

Every value was computed with the model at f9b55fc4dd (main) and at this head. Under the filing exception, the baseline keeps the dependents (Publication 501, Dependent Taxpayer Test); the repeal drops them and keeps the claimable spouse at zero.

Invariants

For every return in Hawaii or Virginia:

1. Dependents only: each exemption output under the repeal equals the baseline output for the same return with its dependents removed (hi_regular_exemptions, hi_disabled_exemptions, hi_exemptions, va_personal_exemption, va_aged_blind_exemption, va_total_exemptions). The same holds for each filer's va_personal_exemption_person.
2. Never above baseline: no output is higher under the repeal than in the baseline. In particular, the repeal never raises a claimable filer's exemption.
3. No VA personal exemption goes to a claimable filer or to a dependent under the repeal.

policyengine_us/tests/core/test_claimable_filer_hi_va_repeal_invariants.py tests these in two ways:

• An exhaustive grid, run for 2021, 2025 and 2026: 240 returns. Each filer is not claimable, claimable but unclaimed, claimed, or within the filing exception. Returns are single or joint; filers are young, aged, or blind and disabled; each return has zero or two dependents.
• Hypothesis batches of 6–12 generated returns.

Invariant 1 is a differential test against an independent path: the baseline formulas run on the return without its dependents. It does not re-implement the reform.

Tests

Each file ran alone in a foreground process, with OMP/OpenBLAS/MKL/vecLib/NumExpr capped at 4 threads and OMP_WAIT_POLICY=PASSIVE.

File	Main formula	This PR
contrib/state_dependent_exemptions/repeal_state_dependent_exemptions.yaml	11 failed, 11 passed	22 passed
core/test_claimable_filer_hi_va_repeal_invariants.py	4 failed	4 passed
baseline/gov/states/hi/tax/income/exemptions/ (4 files, each alone)	not run (baseline unchanged)	10, 2, 10 and 6 passed
baseline/gov/states/va/tax/income/exemptions/ (4 files, each alone)	not run (baseline unchanged)	5, 5, 6 and 2 passed
baseline/gov/states/ok/tax/income/ok_count_exemptions.yaml (also applies the repeal)	not run	7 passed

• Repeal YAML. The 15 new cases are 11 claimable-filer cases and 4 controls. The 11 fail on main with exactly the values in the table. The controls (two baseline twins and two ordinary-filer returns) pass on main and on this head, as they should.
• Mutation checks. Reverting only the VA gate fails all 4 property tests. Dropping only Hawaii's claimable bar on the aged add-on fails all 4 property tests and the 2 aged YAML cases.
• Cost. The property module builds the reformed system once, as a module-scoped fixture released when the module ends. Locally (macOS, load ~9), pytest took 18.5 s, of which fixture setup was 9.4 s. The process took 46.7 s wall including imports, with peak RSS 1.79 GB (/usr/bin/time -l). These are local measurements, not Linux CI figures. The file falls in Rest Python's existing core group (shard 2); no runner, group or process was added.
• Checks. Ruff format and check, and git diff --check, pass. No microsimulation was run.

Impact

• Baseline: no change. The diff touches only override classes inside create_repeal_state_dependent_exemptions(). They are applied only when gov.contrib.repeal_state_dependent_exemptions.in_effect is true, and it defaults to false.
• The contributed repeal: in Hawaii and Virginia, tax units with a claimable head or spouse get lower exemptions than before, which can raise their state income tax under the reform. Units with no claimable filer are unchanged, except for the inconsistent filing_status-input case noted above.
• Microsimulation: not run. No baseline result can move, because the diff is confined to the reform module (policyengine_us/reforms/state_dependent_exemptions/repeal_state_dependent_exemptions.py), its tests and the changelog. The change applies only when the contributed repeal is switched on.

Review

GPT-6.1 Sol (Subfleet 20261019-085035-us10271-rev1, with a shell) approved at c0ceabd55b with no P1–P3 findings. It reproduced:

• all 13 table rows;
• the main-formula failures (YAML 11 failed, property test 4 failed);
• both mutation checks;
• every baseline file.

It also confirmed the conflicting-filing_status edge: two non-claimable filers with filing_status: SINGLE get Hawaii regular exemptions of $1,144 on main, $2,288 at this head, and $2,288 in the baseline.

Follow-ups (not in this PR)

The same reform file has two more overrides that break the "dependents only" invariant on main:

• Delaware. de_personal_credit counts a claimable filer: with $40k wages, the credit is $0 in the baseline and $110 under the repeal. The override also bypasses the non-refundable credit cap and ordering.
• Michigan. A claimable filer within the filing exception keeps a child's exemption under the repeal ($7,300 in both the baseline and the repeal, where the repeal should give $1,500), because mi_exemptions' claimable branch bypasses the overridden mi_personal_exemptions.

#10044 fixes RI and WI. Delaware and Michigan are queued as a separate PR.

axiom: n/a: contributed reform; no statute encodes it (the baseline HI and VA rules are unchanged)

🤖 Generated with Claude Code

[MaxGhenis]

Hand-off to the US + core hub.

• Head: c0ceabd55b0c58604781206c598710fb98b629ce, base main f9b55fc4dd. Marked ready 2026-10-19 ~13:27Z.
• Independent review: GPT-6.1 Sol, with a shell (Subfleet 20261019-085035-us10271-rev1), returned APPROVE at this exact head with no P1–P3 findings and METHODOLOGY_CHOICE: no. Path: ~/reviews/us-hub/reviews-gen/us-10271/review-r1.md.
• Passed locally, each file alone, threads capped at 4:
  • repeal YAML: 22 passed (11 fail on main's formula);
  • core/test_claimable_filer_hi_va_repeal_invariants.py: 4 passed (4 fail on main);
  • HI and VA baseline exemption YAML: 10/2/10/6 and 5/5/6/2 passed;
  • ok_count_exemptions.yaml: 7 passed.
  • Ruff format and check, and git diff --check: clean.
• Pending: PR CI (started when the PR was marked ready).
• Dependencies: none. git merge-tree against Apply the RI, SC, UT, VT, WV and WI rules for a filer who can be claimed as a dependent #10044's head 15a794021e is clean: the YAML cases sit after the Vermont cases, not at the end of the file.
• Holds and decisions: none. This is a contributed-reform bug fix with no baseline change; axiom: n/a.
• Impact: no baseline change. Microsimulation of the reform's change is the hub's.
• Follow-ups owed: a separate PR for the Delaware and Michigan overrides in the same file. Both are confirmed by execution, and the evidence is listed in the body. It is proposed as a session card ("Fix DE and MI overrides in the dependent-exemption repeal"). Evidence: ~/reviews/us-hub/fixes/hi-va-repeal-claimable-10271/.

[DTrim99]

PR Review: Keep claimable filers' own exemptions under the HI and VA dependent-exemption repeal

PR #10271 (author: MaxGhenis). Head c0ceabd55b. The branch is 2 commits ahead of main and 42 behind. First review.

Summary

The contributed reform gov.contrib.repeal_state_dependent_exemptions overrides hi_regular_exemptions and va_personal_exemption_person. On main, those overrides counted every head and spouse: head_spouse_count in Hawaii and is_tax_unit_head_or_spouse in Virginia. As a result, a filer whom another taxpayer can claim got the $1,144 Hawaii exemption (plus the aged add-on) or the $930 Virginia exemption under the repeal. The baseline gives that filer $0.

So a "dependent exemption repeal" raised the exemptions of claimable filers. The PR swaps each override's filer count for the baseline's own gate and drops only the dependent term:

• Hawaii: head_spouse_count_not_dependent_elsewhere, plus aged & head_or_spouse & ~claimable.
• Virginia: filer & ~claimable.

What the change gets right:

• Intent. "Repeal dependent exemptions" should remove only the exemptions claimed for dependents. A claimable filer's own exemption should stay as baseline law sets it, which is zero in both states. The PR does exactly that.
• Law. It is right under HRS 235-54(a), Va. Code 58.1-322.03(2)(a) and IRC 151(d)(2).
• Code. The change is vectorized and stays inside the reform's override classes. No baseline file changes.
• Tests. Every new reformed YAML case activates the reform through the top-level reforms: key. The new Python invariant test builds the reformed system directly.
• CI. Green: 27 pass, 3 skipping. I hand-checked all 15 new YAML expectations, and they are correct.
• Ratchet. The source-conflict ratchet from Tie source-conflict markers to the register shared with Axiom #10287 isn't affected.

Open items: no critical items, no should-address items, and three suggestions.

Critical (must fix)

None.

Should address

None.

Suggestions

1. The title and changelog wording read backwards.

• Where. changelog.d/fix-hi-va-repeal-claimable.fixed.md:1, and the PR title.
• What is wrong. Both open with "Keep a claimable filer's own exemptions under the ... repeal". Most readers will take that to mean the claimable filer keeps a (nonzero) exemption. The change does the opposite: it stops the repeal from granting an exemption the baseline denies. The clause after the colon states the effect correctly. But the changelog line is what users see, and the lead clause invites the misreading.
• Fix. Something like: "Stop the Hawaii and Virginia dependent-exemption repeal from granting a regular, aged or personal exemption to a filer whom another taxpayer can claim (HRS 235-54(a); Va. Code 58.1-322.03(2)(a))."

2. The new test module imports its strategies from another test module.

• Where. policyengine_us/tests/core/test_claimable_filer_hi_va_repeal_invariants.py:35-38.
• What is wrong. The file imports adults and dependents from test_dependent_elsewhere_head_spouse_swap_invariance.py. It works today, and pytest doesn't re-collect any tests because only the strategies are imported. But if someone renames that module or changes those strategies (for example, the income ranges or dependent ages), this test changes or breaks with no sign in its own file.
• Fix (optional). Move the two strategies into a small shared helper module under tests/core/, or define them locally.

3. Track the deferred Delaware and Michigan overrides in an issue.

• The deferral. The PR body documents two more overrides in the same reform file that break the "dependents only" invariant on main:
  • de_personal_credit counts a claimable filer, and it bypasses the credit cap and ordering;
  • for mi_exemptions, a claimable filer within the filing exception keeps a child's exemption.
• What is missing. These are queued only as an internal session card. I found no public issue or PR for them.
• Fix. Open an issue that carries the execution evidence from the body, so the follow-up is tracked publicly. (I spot-checked the other head_spouse_count / tax_unit_size overrides: MD, NE, WV and KS. Their baselines don't zero a claimable filer either, so "two more" looks accurate.)

Verified correct

The law.

• HRS 235-54(a) (HRS ch. 235, file page 45): "In the case of an individual with respect to whom an exemption under this section is allowable to another taxpayer ..., the personal exemption amount applicable to such individual under this subsection for such individual's taxable year shall be zero."
  • The additional exemption for "the taxpayer or the taxpayer's spouse who is sixty-five years of age or older" sits in the same subsection, so it is zero too.
  • The 2025 N-11 instructions agree (file page 9): "If you can be claimed as a dependent on another person's tax return, do not fill in the ovals on lines 6a and 6b."
  • The HRS 235-54(c) $7,000 disability exemption has no such bar. The PR leaves hi_disabled_exemptions alone, which is correct.
• Va. Code 58.1-322.03(2)(a): "A deduction in the amount of $930 for each personal exemption allowable to the taxpayer for federal income tax purposes."
  • IRC 151(d)(2): for an individual whom another taxpayer can claim, "the exemption amount applicable to such individual for such individual's taxable year shall be zero."
  • 58.1-322.03(2)(b) gives $800 to "each blind or aged taxpayer" with no dependency bar. The PR leaves va_aged_blind_exemption_person alone, which is correct.
  • Context, outside this PR. The 2025 Form 760 instructions (file page 15) and Virginia Tax's Exemptions page ("Each filer is allowed one personal exemption") don't address a claimable filer's own box. The baseline's reading rests on the statute and IRC 151(d)(2). This PR doesn't change that reading. And if the baseline is ever revisited, the new invariant test will force the reform to follow it.

The code.

• Hawaii override.
  • It equals the baseline hi_regular_exemptions minus the dependents. The baseline exemptions_count returns own + dependents when a filer is claimable, and tax_unit_size otherwise. With the dependents removed, both reduce to head_spouse_count_not_dependent_elsewhere.
  • The aged term is identical to the baseline's (aged & head_or_spouse & ~claimable).
  • The reference tuple, including the HRS #page=45 anchor, matches the baseline and shows the "shall be zero" sentence.
• Virginia override. It equals the baseline va_personal_exemption_person without its dependent & ~dependent_filer term.
• The filing_status edge. The filing_status formula returns JOINT exactly when tax_unit_married is true. So with no claimable filer, the old and new Hawaii counts agree unless a filing_status input contradicts the unit's members. In that case the new count follows the baseline, as the body says.
• Reform activation is unchanged. It still runs through create_repeal_state_dependent_exemptions_reform (in_effect looked up over a five-year window, bypass for tests) and the registration in reforms.py. With in_effect false, baseline HI and VA outputs can't change: the diff touches only the reform file, the two test files and the changelog.
• Vectorization. Only array operators: &, ~ and tax_unit.sum. There are no state-gated enum or int keys, so the Add New Hampshire state supplementation for 2026 #10156 pattern doesn't apply.
• Changelog. A top-level changelog.d/<name>.fixed.md, a sensible type.
• No partner test files are touched.

Tests activate the reform.

• All 13 reformed YAML cases carry reforms: policyengine_us.reforms.state_dependent_exemptions.repeal_state_dependent_exemptions.repeal_state_dependent_exemptions. That is the bypass-built Reform, so it always applies. The two baseline twins have no reforms: key and set in_effect false.
• The YAML uses only the two override values (true and false) that the file already used, so the PR adds no extra system copies.
• The Python test builds CountryTaxBenefitSystem(reform=repeal_state_dependent_exemptions) once per module.
• The Python test is worth its cost. It runs a 240-return grid for 2021, 2025 and 2026, plus derandomized Hypothesis batches. It compares the repeal against the baseline run on the same returns with their dependents removed, which is an independent path, not a re-implementation. That catches drift if the baseline HI or VA formulas change later.
• Grid count. 2 states × 2 dependent options × (3 single-filer sets × 4 claim states + 3 couple sets × 16) = 240.
• Cost on Linux CI (Rest Python shard 2, core group): 20.75 s fixture setup and about 5.8 s of calls. The PR body's local figures (18.5 s; 1.79 GB peak RSS) are labelled as local. No runner or group was added.

Hand-checked expectations (all 15 new YAML cases, 2025). HI base $1,144 and aged threshold 65; VA personal $930 and aged/blind $800.

#	Case	Repeal (this PR)	Arithmetic	Baseline / main repeal
1	HI single, 22, claimable, baseline twin	regular 0, total 0	own 0 + deps 0 + aged 0	–
2	HI single, 22, claimable, unclaimed	0, 0	not-dependent-elsewhere count 0	main 1,144
3	HI joint 40 / 22 (spouse claimable), 2 children	1,144	1 own; dependents dropped	baseline 1,144 (152(b)(1)); main 2,288
4	HI joint, claimable head in filing exception, 2 children	1,144	1 own	baseline (1+2) × 1,144 = 3,432
5	HI joint, both claimable	0, 0	0 own, 0 aged	main 2,288
6	HI single, 66, claimable	0	aged term masked by ~claimable	main (1+1) × 1,144 = 2,288
7	HI joint 67 (claimable) / 70	2,288	(1 own + 1 aged) × 1,144	baseline 2,288; main 4,576
8	HI joint 70 / 68 + child (control)	4,576	(2 + 2 aged) × 1,144	baseline 5 × 1,144 = 5,720
9	VA single, 22, claimable, baseline twin	personal 0, total 0	filer & ~claimable = 0; not aged	–
10	VA single, 22, claimable, unclaimed	0, 0	same	main 930
11	VA joint 40 / 22 (spouse claimable), 2 children	930; per person [930, 0, 0, 0]	head only	baseline 930; main 1,860
12	VA joint, claimable head in filing exception, 2 children	930; [0, 930, 0, 0]	spouse only	baseline 3 × 930 = 2,790
13	VA joint, both claimable	0, total 0	none eligible; not aged	main 1,860
14	VA single, 66, claimable	personal 0, aged/blind 800, total 800	(2)(b) has no bar	main 930 + 800 = 1,730
15	VA joint 40 / 38, 2 children (control)	1,860; [930, 930, 0, 0]	2 × 930	baseline 4 × 930 = 3,720

• CI ran them all. The contrib other-shard-1 job ran the repeal YAML at this head: 22 passed (the 7 existing cases plus these 15).
• The PR body's 13-row table matches this arithmetic for both the baseline and main columns.

Axiom line assessment

axiom: n/a: contributed reform; no statute encodes it (the baseline HI and VA rules are unchanged). Valid.

• The diff changes only an override inside a contributed (hypothetical) reform, plus its tests and changelog.
• No enacted-law provision changes. No baseline variable or parameter is touched, so there is no rulespec-us module to bring to parity.

CI status

• Result. gh pr checks 10271, re-checked at the end of the review: 27 pass and 3 skipping, all at this head (run 39260154092). The 3 skips are the "Baseline / Contrib / Python tests" entries from an earlier run (39254402520). Nothing is pending and nothing failed.
• The ratchet.
  • The run started 2026-10-19, before Tie source-conflict markers to the register shared with Axiom #10287 (the source-conflict ratchet) merged on 2026-10-20.
  • I checked the ratchet statically against the merged tree. The only scanned file the PR touches is policyengine_us/reforms/state_dependent_exemptions/repeal_state_dependent_exemptions.py. It has 0 comment lines matching the ratchet's conflict pattern at the merge base and 0 at this head, and it is not in KNOWN_UNMARKED.
  • main has not changed that file since the merge base. Test directories aren't scanned.
  • So the merged tree passes the ratchet.

Branch status

• Position. 42 behind main, 2 ahead (merge base f9b55fc4dd).
• Merge with main. git merge-tree --write-tree PolicyEngine/main c0ceabd55b is clean (tree c05adc0ae2). Since the merge base, main has not touched the reform file, the contrib YAML, the HI or VA exemption variables, or the imported swap-invariance test module.
• Sibling note. The PR also merges cleanly with Apply the RI, SC, UT, VT, WV and WI rules for a filer who can be claimed as a dependent #10044's current head fc6bd175bb (RI and WI overrides in the same reform file).

Verdict: APPROVE

[DTrim99]

The fix is right. The HI and VA repeal overrides now use the baseline filer gate minus dependents, so a claimable filer gets the same $0 own exemption as under baseline law (HRS 235-54(a); Va. Code 58.1-322.03(2)(a) with IRC 151(d)(2)). Baseline is unchanged, all 13 reformed cases use reforms:, and the merged tree passes the source-conflict ratchet. Three suggestions, including rewording the title, which reads as if the filer gains an exemption. Full report in the comment above.

[MaxGhenis]

US + core hub merge audit, us#10271 at c0ceabd55b0c58604781206c598710fb98b629ce (squash): Keep claimable filers' own exemptions under the HI and VA dependent-exemption repeal.

• Review: GPT-6.1 Sol review r1 (workspace-write, shell-run tests) APPROVE at c0ceabd55b, reviews-gen/us-10271/review-r1.md, methodology no. DTrim99 APPROVED this head on 2026-10-21.
• 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: none on the baseline: the diff is confined to the contributed repeal's reform module, its tests and the changelog, so no baseline result moves. The reform's own change (claimable HI and VA filers keep the baseline's $0 own exemption) is covered by the YAML and property tests; no microsimulation was run.
• Axiom: n/a: contributed reform; no statute encodes it (the baseline HI and VA rules are unchanged)