Give each spouse their own above-the-line deductions in per-person AGI

[MaxGhenis]

Summary

This PR attributes each spouse's federal above-the-line deductions to the person who owns them. adjusted_gross_income_person currently splits most deductions equally on a joint return, so one spouse's IRA deduction or business loss can lower the other spouse's state AGI. The new person-level allocator preserves the federal tax-unit totals while changing the person amounts read by state taxes, benefits and withholding estimates.

• above_the_line_deductions_person combines each person's deduction amounts. IRA deductions, early withdrawal penalties, student loan interest, adoption assistance and U.S. savings bonds for education follow the person who has them.
• Tax-unit deductions with person amounts—self-employment tax, self-employed health insurance and retirement plans, alimony, business and capital losses, educator expenses and HSA deductions—follow those amounts. A directly supplied tax-unit amount is reconciled using the spouses' own amounts, with an equal split when neither has an amount.
• Tax-unit-only deductions—tuition and fees, domestic production, possession and Puerto Rico income—remain equally divided. HSA deductions supplied only at tax-unit level also use that fallback.
• loss_ald_person divides the allowed business loss by each spouse's own business, farm, rental, estate and partnership losses. A tax-unit-only Form 4797 loss counts equally. The capital portion follows each spouse's own capital losses, including losses allowed against gains and the net capital-loss limit. Main's limited_business_loss supplies the federal Section 461(l) calculation; limited_capital_loss_person and alimony_expense_ald_person provide the other person amounts.
• filer_share computes a head's or spouse's share, with an equal fallback and zero for dependents. A dependent's above_the_line_deductions_person keeps their own person-level deductions for their own return, excluding filer amounts recorded on dependents. Their adjusted_gross_income_person, loss_ald_person and limited_capital_loss_person remain zero.
• adjusted_gross_income_person becomes irs_gross_income − above_the_line_deductions_person. The head's and spouse's amounts still sum to federal adjusted_gross_income.
• Montana through 2023 assigns its non-loss adjustments to their owners. Its business and capital losses retain the existing equal split through mt_loss_ald_reallocation; Montana's historical loss elections and separate Section 461(l) threshold remain follow-ups. From 2024, mt_agi_indiv, mt_applicable_ald_deductions and mt_loss_ald_reallocation apply only while separate filing on the same return is eligible, so they produce no spouse allocation after that filing arrangement ended.
• DC's new dc_separate_capital_loss_adjustment replaces the capital portion actually carried in each spouse's federal AGI with their own capital losses, limited to their own gains and distributions plus $1,500. The separate taxable-income and deduction-allocation paths use it. The joint taxable-income path preserves the joint federal deduction. dc_self_employment_loss_addition uses min(limited_business_loss, loss_ald) for the business component and retains main's exclusion of dependent losses. The business component agrees with main when loss_ald is computed; directly supplied totals use business losses first.

The Montana joint standard and medical deduction fixes already landed in #9903. This PR uses main's formulas and tests for those deductions, including the federal medical expense deduction from 2024.

The merge also retains main's per-educator cap through educator_expense_ald_person (#9833) and person HSA input health_savings_account_ald_person (#9865). The spouse allocator uses those person amounts and reconciles them to the return total; HSA is equally divided only when no filer account-holder amount is available.

Consumers and modeling choices

Max approved the state-specific loss treatment in decision d1014.

Consumer	Treatment and basis
Kentucky combined-separate AGI	Each spouse's own deductions; joint limited losses allocated proportionally to their own losses. KRS 141.180(3) and Form 740 require own attribution without specifying a joint-limit ratio.
Delaware de_pre_exclusions_agi	Own income, losses and deductions, using amounts allowed on the joint federal return. 30 Del. C. 1162(b)(1) and PIT-RES line 1; proportional allocation is the d1014 default.
DC separate filing on the same return	Each spouse's own capital-loss deduction with the $1,500 ceiling; business losses proportional. Schedule S Calculation J and D-40 line c.
Virginia va_agi_person and Spouse Tax Adjustment	Spouse columns reconcile to joint federal AGI. Proportional joint-limit allocation follows d1014; Virginia additions are prorated by positive AGI.
Montana through 2023	Non-loss adjustments go to their owners under ARM 42.15.206(1) and former MCA 15-30-2110(8)–(9). Business and capital losses retain the existing equal split. Student loan interest follows the payer; the historical instructions also permit equal or AGI-proportional allocation.
Ohio joint filing credit	Each spouse's qualifying income reflects their own deductions; proportional losses follow d1014. R.C. 5747.05(E) requires $500 of qualifying income without specifying a joint-loss ratio.
West Virginia senior/disability deduction	Each eligible person's attributed AGI on a joint state return, under W. Va. Code 11-21-12(c) and Schedule M. The separate-state-return recomputation rule has a different scope.
Illinois Family Planning, Minnesota renters' credit, Utah homeowner/renter relief and state withholding estimates	Read per-person AGI. Michigan and Pennsylvania withholding estimates now floor their amounts at zero, matching the other states' treatment.

Proportional allocation where state law leaves the joint-limit ratio unspecified is a modeling choice, using each spouse's own losses as the ownership measure. It gives all of a sole spouse's loss to that spouse. Tax-unit-only inputs retain an equal fallback; for HSA, 26 USC 223(b)(5) provides equal allocation unless spouses agree otherwise.

Invariants and coverage

The spouse property tests cover KY, MT, DE, VA, DC, OH and WV in 2020, 2023 and 2026:

• Person above-the-line, loss, capital-loss and computed alimony amounts reconcile to their tax-unit totals, including directly supplied deduction totals. Person AGIs reconcile to federal AGI.
• Independent numpy calculations check the Section 461(l) and capital-loss limits, each educator's separate cap, and the person allocator, netting capital losses against gains and distributions first.
• Own allowed deductions stay with their owner. Monotonicity and independence checks account for shared IRA compensation, retirement-plan coverage and joint MAGI effects, as well as student-loan shared limits and MAGI interactions. Removing spouse inputs can free IRA compensation or change coverage and phase-outs, increasing the head's own allowed deduction. Removing a spouse's losses can increase the head's share of a joint limited loss. AGI independence also excludes Social Security and unemployment amounts taxed according to joint AGI. These checks exclude directly supplied tax-unit deduction totals.
• Dependents receive no filer AGI or loss allocation. Deductions are nonnegative unless a negative total is supplied directly.
• DC separate capital deductions equal each spouse's own loss within their gains plus $1,500. Computed joint totals do not produce a greater combined separate deduction; directly supplied loss_ald can require a negative adjustment.
• Montana allocates nothing through the gated variables from 2024.
• Dependent-deduction regressions and hash-seed determinism have separate Python coverage. The deduction-coverage audit reads the already initialized baseline system's registry and parameter histories.

YAML cases cover sole and joint loss owners, deductions supplied directly, DC capital-loss limits and mixed business/capital losses, Montana's pre-2024 and post-2023 paths, state AGI consumers, and withholding floors. The r8 direct DC example (loss_ald: 22000, business loss $20,000, capital loss $5,000) expects a business add-back of $8,000, capital adjustment of $500 and separate taxable incomes of $17,300/$40,000.

The new HSA reconciliation case supplies person amounts of $1,000/$3,000 and a return deduction of $6,000, expecting person deductions of $1,500/$4,500. Two 2020 IRA compensation regressions preserve ownership while showing the shared-limit interaction: $1,000/$3,000 of wages and $3,000 of contributions each yield deductions of $1,000/$3,000; removing the higher earner's contributions frees compensation for the lower earner, yielding deductions of $3,000/$0 and person AGIs of −$2,000/$3,000.

The determinism test retains eight unit cases and six independent hash-seed interpreters, with one complete US household simulation as the oracle. Each seed child uses actual core holders and projectors plus five production formulas (above_the_line_deductions, above_the_line_deductions_person, loss_ald_person, adjusted_gross_income_person and student_loan_interest_ald_magi), using live parameter lists and explicit dependency values captured from that oracle. Values, dtypes and exact bytes must agree. The children focus on summation; default country construction and other deduction eligibility are evaluated once by the parent.

Tests

Completed sequential Python runs at a1a10fc21ff1969e2bae9a929368907f84a5122a:

File	Passed	Elapsed	Peak memory
test_spouse_above_the_line_deductions_invariants.py (new)	7	193.76 s	961.8 MiB
test_dependent_above_the_line_deductions_invariants.py	9	724.37 s	2230.3 MiB

The later determinism-only change and conflict-free main merge do not change these test files or the spouse allocator. make format passes.

Required YAML validation remains blocked on local runtime conditions: 132 cases in 19 explicit baseline files, plus 8 cases in partners/my_friend_ben/2024.yaml and partners/analytics_coverage/signatures/il.yaml. The first permitted 12-file baseline batch was interrupted in startup without reaching a case; the other baseline batch and partner files were not run. Shared-machine load averages exceeded 325. Python files and partner files run individually, with only one test process active. Partner expectations remain unchanged; pending runs cannot establish whether any partner output moves.

Both the original merged-head determinism test and its revised isolated seed fixture passed eight unit cases before a child exceeded the 180-second timeout on the busy shared machine. An isolated diagnostic of the revised fixture later completed with the expected float32 values and bytes. These attempts are preserved as failed attempts, not passing tests. A further retry was interrupted after roughly half an hour in startup without test output. Complete determinism validation remains blocked on local runtime conditions.

Dependencies and follow-ups

• Allocate Delaware itemized deductions by payer and fix the 529 subtraction limits #9878 has merged: Delaware's unallocated itemized deductions use positive-income proration.
• Allocate Virginia's 529 deduction to the owner of record and deduct it after VAGI #9875 remains an open prerequisite: Virginia's 529 deduction must move to the account owner and stop prorating by possibly negative AGI.
• Net household members' Medicaid MAGI before the zero floor, and give each spouse their own deductions #9860 is stacked on this PR and owned by the hub. It changes Medicaid to each spouse's own deductions together with the household income floor; Give each spouse their own above-the-line deductions in per-person AGI #9890 must land first. Medicaid's person AGI remains unchanged here.
• West Virginia separate state returns and Virginia filing status 3 after a joint federal return require hypothetical separate federal AGI, including the $1,500 capital-loss ceiling and a separate Section 461(l) threshold. Modeling those separate-state elections and recomputations remains a follow-up.
• Montana through 2023 still needs ARM 42.15.206(3) capital-loss elections, a recomputed single-filer Section 461(l) threshold, and review of the child/dependent-care phase-out's AGI base. That deduction ended after 2023.

axiom: TheAxiomFoundation/rulespec-us#1517 queued

Hub microsimulation impact (current head)

A real microsimulation on the default dataset, main at the merge base 4fbf8476b5 against this PR's head f65aab3e3d, for 2026. The runs went one at a time under the shared heavy-job lock. Sums are weighted.

Output	2026 change	2026 records changed (weighted)
adjusted_gross_income_person	-1,828	18,366 (41,902,731)
state_income_tax	+12,715	176 (162,012)
ky_income_tax	+14	3 (0)
de_income_tax	-826,766	55 (12,643)
dc_income_tax	+683,183	20 (6,118)
va_income_tax	+276,107	32 (95,851)
oh_income_tax	+10	5 (0)
wv_income_tax	-13,619	3 (5,351)
mi_income_tax	+0	0 (0)
pa_income_tax	+0	0 (0)
mt_income_tax	-10	2 (1,631)
income_tax	+25.45M	394 (722,081)
household_net_income	-25.46M	533 (858,483)
household_benefits	+0	0 (0)

Where the federal change comes from. Federal AGI does not change. salt_deduction takes reported_salt, which takes state_and_local_sales_or_income_tax, the larger of sales tax and state_withheld_income_tax. All 42 state *_withheld_income_tax estimates on main read adjusted_gross_income_person, so moving a deduction to the spouse who owns it changes the estimated state withholding that is itemized as SALT, and with it federal taxable income. A hub breakdown on the fixed 10% household sample (a check only; the figures above are the full-file run) places the whole income_tax change in the itemized SALT deduction. AGI, the standard deduction, the QBI deduction, AMT, NIIT, the additional Medicare tax, the EITC, CDCC, saver's credit, PTC, AOC, the elderly credit and taxable Social Security do not change on that sample; the CTC moves only between its refundable and non-refundable parts in 5 records. Variables this PR adds (loss_ald_person, above_the_line_deductions_person) have no value on the base side, so the table leaves them out.

Comparison files are in fixes/9890-impact/ on the hub host.

[MaxGhenis]

US + core hub merge audit, us#9890 at f65aab3e3d803a3eaf5c8134e76e66808db143ca (squash): Give each spouse their own above-the-line deductions in per-person AGI.

• Review: GPT-6.1 Sol review r10 APPROVE at f65aab3 (reviews-gen/us-9890/review-r10.md): KY double loss allocation fixed; no open findings. Impact re-measured at this head (full file, 2026); federal income_tax +25.45M traced to the SALT deduction through state withholding estimates (10% screen breakdown, fixes/9890-impact/compare.screen.2026.json).
• 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 (hub full-file microsimulation, merge base 4fbf8476b5 against this head, in the description): 2026: adjusted_gross_income_person -1,828 (18,366 records); state_income_tax +12,715 (176 records); ky_income_tax +14 (3 records); de_income_tax -826,766 (55 records); dc_income_tax +683,183 (20 records); va_income_tax +276,107 (32 records); wv_income_tax -13,619 (3 records); income_tax +25.45M (394 records); household_net_income -25.46M (533 records).
• Order: Lands before Net household members' Medicaid MAGI before the zero floor, and give each spouse their own deductions #9860 (Net household members' Medicaid MAGI before the zero floor, and give each spouse their own deductions #9860 has after: us#9890).
• Axiom: Encode spouse attribution of federal adjustments in state per-spouse income computations TheAxiomFoundation/rulespec-us#1517 queued