Subtract Mississippi's HSA deduction once, from the account holder

[MaxGhenis]

Mississippi subtracted a tax unit's health savings account deduction from every member's income. ms_agi_adjustments is a Person variable whose adds list held the TaxUnit variable health_savings_account_ald, and policyengine-core maps a group value onto a person target by copying it to each member (Simulation.map_result). A couple therefore subtracted the deduction twice, and each dependent with income once more (their Mississippi net income goes to the head's AGI). Reported in #9801's hand-off as a pre-existing follow-up.

The law

• Whose deduction. Miss. Code Ann. § 71-9-11(3) excludes the principal contributed to and the interest earned on a health savings account "from the taxable gross income of the account holder under Section 27-7-15" (H.B. 1213 (2005) § 2). The account holder is the individual "for whose benefit" the account is established (§ 71-9-3(b)).
• Where it goes. Form 80-105 reports adjustments in Column A (Taxpayer) and Column B (Spouse), with an HSA deduction line (2022 form, line 59). The 2025 instructions say a combined return puts each spouse's income in their own column and separates incomes "throughout the return" (pp. 5–6), and the HSA line takes "the amount deposited into a health savings account" (p. 13).
• Dependents. A dependent has no HSA deduction on this return, and none of their own under 26 USC 223(b)(6).

The change

• ms_health_savings_account_adjustment (new, Person) attributes the tax unit's health_savings_account_ald to the head and spouse with Keep dependents' above-the-line deductions off the filer's AGI #9801's person_share_of_tax_unit_amount. Each gets their own health_savings_account_ald_person, scaled so the shares sum to the tax unit's deduction. A dependent's share is zero, even if the dependent has their own amount: a dependent who holds an account reports it on their own return.
• Default without per-person amounts. The head takes the deduction, as for the other tax-unit amounts that helper reconciles. The dataset has no per-person HSA amounts, so every couple in the microsimulation uses this default. Limitation: a head whose Mississippi income is below the deduction loses the excess to ms_agi's zero floor. A YAML case pins this; a per-person amount on the spouse avoids it. Without the account holder's identity, a 50/50 split (the federal per-person AGI convention) is no better grounded in law.
• health_savings_account_ald_person (new Person input) is each account holder's own deduction. It only splits the deduction: health_savings_account_ald stays the tax-unit input the dataset supplies. Giving that variable a formula would stop core from uprating the 2024 dataset value, because core runs a formula before it uprates.
• The Mississippi adjustments list uses ms_health_savings_account_adjustment in place of health_savings_account_ald.

Tests

• YAML. ms_health_savings_account_adjustment.yaml has 9 cases:
  • single;
  • joint without per-person amounts;
  • the spouse's account;
  • two account holders;
  • per-person amounts that differ from the tax-unit total;
  • a dependent with wages and their own account;
  • a married person filing a separate federal return;
  • separate-return columns, with tax computed by hand;
  • the head-default limitation.
• Property test. tests/core/test_ms_health_savings_account_adjustment_invariants.py runs three Hypothesis batches of 10–40 Mississippi tax units, plus a seeded population of 200. Units are single, married filing separately, head of household, or joint with and without dependents. Dependents may have their own HSA amounts.
• Fails on main, passes here. On main (bab1a47c60), the joint case gives ms_agi_adjustments [3,000, 3,000] where [3,000, 0] is expected; the dependent case gives [3,000, 3,000, 3,000]. A main-compatible probe of the property test found the deduction counted 2 to 4 times in 120 of 200 seeded units, and total Mississippi AGI reduced by more than the deduction in 89 units (by up to $17,890). On this branch: 0 and 0. All 457 Mississippi YAML cases and test_ald_determinism.py (9/9) pass.
• Cost. The property test takes 26 s and 1.0 GB peak locally (macOS). It runs in the sequential rest-core Python group (make test-other-python-rest-core), with no new jobs or runners. GitHub Actions has had every recent PR-workflow run queued, so there is no fresh Linux timing or memory report for that group to compare against.

Invariants (for every input)

1. Conservation. The members' ms_health_savings_account_adjustment sum to the tax unit's health_savings_account_ald, and that deduction appears exactly once in the members' ms_agi_adjustments.
2. Bounds. No share is negative or above the tax unit's deduction, and a dependent's is zero.
3. Attribution. When the head's and spouse's own amounts sum to the deduction, each gets their own; without any, the head gets all of it.
4. Monotonicity. The deduction lowers the members' total Mississippi AGI by at least zero and at most the deduction.

Impact

Real microsimulations on the default dataset (populace_us_2024.h5@populace-us-2024-spm-20260909), one run per checkout and year under the shared heavy lock (lockrun.sh, run.py, compare.py in ~/reviews/pe-us-merge-backlog-2026-10-02/impact/), compared record by record; enum arrays are not compared. Outputs: ~/reviews/pe-us-ms-hsa-qbid-2026-10-06/microsim/. Base: main bab1a47c60; branch head d5e8301b9c.

	2025	2026
Mississippi adjustments (ms_agi_adjustments, persons)	−$352.2M (31 records, 36,875 people, all down)	−$371.4M (31 records, 37,139 people)
Mississippi AGI (ms_agi)	+$89.3M (8 records, 14,845 people, all up)	+$94.2M
Mississippi income tax	+$3.87M (6 records, 9,219 tax units, all up; largest +$506)	+$3.71M (9,285 units; largest +$495)
Federal income tax	0	0
health_savings_account_ald	unchanged	unchanged

The removed adjustments are the extra copies of the deduction on spouses and dependents. Most of those people had too little Mississippi income for the copy to matter, because ms_agi floors at zero, so AGI rises by much less than the adjustments fall. No household's Mississippi tax falls.

Axiom

axiom: TheAxiomFoundation/rulespec-us#1509 queued

Mississippi's HSA exclusion is not encoded, and Miss. Code Title 71 ch. 9 is not in the pinned corpus. Ingesting it is the encoder blocker; signed encoder runs also need Max's approval.

🤖 Generated with Claude Code

[MaxGhenis]

Hand-off to the US + core hub. Head d5e8301b9c. Only CI remains; it is queued on GitHub's side.

Review (Opus 5.5, subfleet run --task review --tier standard -m opus), in ~/reviews/pe-us-ms-hsa-qbid-2026-10-06/review/:

Round	Head	Verdict	Report
r1	a0b445c	APPROVE, with P2/P3 notes	review-ms-hsa-r1.md
r2 (delta)	d5e8301	APPROVE	review-ms-hsa-r2.md

• The r1 findings were fixed in d5e8301b9c.
• r1's P2 asked to keep the head default and document its limitation. That is done in the PR body, with a YAML case pinning it.
• The r2 P3s are a PR-body count (fixed) and two test nits (left as they are): max_examples and an unused dependent keyword.
• Neither reviewer had a shell, so their checks were by reading. My runs are below.

Passed locally.

• The new YAML file (9/9).
• All 457 Mississippi YAML cases, at the first commit.
• test_ald_determinism.py (9/9).
• The new property test (26 s, 1.0 GB).
• On main, the new cases fail and a probe of the property test shows the double count. Evidence is in ~/reviews/pe-us-ms-hsa-qbid-2026-10-06/before-after/.

Microsim (default dataset, lockrun.sh): Mississippi income tax rises $3.87M in TY2025 and $3.71M in TY2026; federal income tax is unchanged. Details are in the PR body; outputs are in .../microsim/.

Axiom: TheAxiomFoundation/rulespec-us#1509, queued (pe-parity). The blocker is that Miss. Code Title 71 ch. 9 is not in the pinned corpus.

No Max decision needed. This is a bug fix that moves aggregates.

Dependencies: none. It touches only Mississippi files and the new health_savings_account_ald_person input. The open branch ms-joint-combine-negative-column (another session) also touches Mississippi; whichever lands second should rebase.

Follow-ups (not in this PR):

• Iowa and Mississippi move a dependent's net income onto the head's state return #9887: whether Mississippi (and Iowa) should move dependents' net income onto the head's return.
• A chip: audit other person-level adds that copy tax-unit amounts (Montana casualty loss, Ohio medical, West Virginia U.S. government interest, and others). Branch fix-mt-casualty-loss-misc-deduction may already cover Montana.

[MaxGhenis]

US + core hub merge audit, us#9865 at d5e8301b9ce34ff38a87fe8002bc0900e7d04ac3 (squash)

• Reviews: independent Opus 5.5 (Subfleet): r1 APPROVE with P2/P3 notes at a0b445c, then delta r2 APPROVE at this head. r1's P2 (keep the head default) is documented and pinned by a YAML case.
• Hand-off: from local_211e84d0. This is a bug fix that moves aggregates, so no Max decision is needed.
• Gates checked live at merge:
  • gh pr checks exits 0 (35 pass);
  • MERGEABLE, and merges clean with current main;
  • not a draft;
  • no CHANGES_REQUESTED review;
  • head pinned.
• Impact: a real microsim on the default dataset. Mississippi income tax rises $3.87M in TY2025 and $3.71M in TY2026, because the HSA deduction is no longer subtracted twice. Federal income tax is unchanged.
• Follow-ups: Iowa and Mississippi move a dependent's net income onto the head's state return #9887 (Mississippi and Iowa dependent-income transfer); the person-level adds audit chip.
• --admin: used only because a required approving review is the sole block.