Net household members' Medicaid MAGI before the zero floor, and give each spouse their own deductions

[MaxGhenis]

Summary

Medicaid household income now sums members' signed MAGI before applying one zero floor to the household total. Each person's Medicaid AGI reads the reviewed above_the_line_deductions_person allocator now on main. With one spouse earning $30,000 and the other losing $20,000, main's per-person floor counts $20,000; this PR counts $10,000 = max(0, $30,000 - $20,000).

The PR also preserves its separate-spouse income membership correction, the conservative positive-income fallback for family-wide relatives without parent links, and the Medicaid dependent filing-requirement correction: a business loss no longer turns other income into unearned income, and 401(k) deferrals do not hide unearned income.

Reconciliation after #9890 landed

Merged main 654a0c9b54a57e6422f4d0418d0d039a3bb3537c into approved head 6c5f5cddd8ad7cb43c21cd84ff281d382e9ecee7 with history-preserving merge commit 77e92b20487a36815a09407193fbec50542b96eb. #9890 landed as 7e9ce42c1856d4b759e55f6f4fca3ba007b9bf2d.

Both branches carried the same concept: own-spouse ALDs, filer/dependent attribution, business/capital loss allocation, HSA ownership reconciliation and equal fallback, filer_share, per-person federal AGI, D.C. capital/business loss adjustments, Montana ALD allocation, MI/PA withholding floors and Virginia additions allocation. Those implementation changes and their tests now belong to main; every model overlap keeps main's version. The duplicate ALD changelog fragment was removed because its provision was already released. loss_ald keeps main's declarative sum, which has the same components and order as the old PR formula.

The remaining delta is one Medicaid changelog, five Medicaid variables, four Medicaid YAML files and the Medicaid household invariant file. The ALD determinism fixture repair is no longer in the delta: #10221 fixed the same fixture on main, and a second merge of main at 121d534b6b takes main's version (none of the probed formulas read unemployment_compensation directly; test_ald_determinism.py 9 passed, the spouse-ALD invariants 7 and the Medicaid household invariants 5 on that tree). No IRS or state formula/test change remains in the PR delta. No additional non-Medicaid fix from #9860 was needed.

Behavioral differences explicitly retained from main

These are static formula comparisons of the old PR with fetched main, not new microsimulation results. Dollar amounts below use only the stated incomes/deductions. The SE examples explicitly supply zero self_employment_tax_ald and zero person equivalents so the differing loss formula is isolated.

Concrete household/input	Old #9860	Main retained by this merge
2025 single age 45, ordinary SE +$10,000, SSTB SE -$5,000	Nets categories first: loss $0, gross $10,000, AGI $10,000. Reversing signs to -$10,000/+$5,000 gives loss $5,000, AGI $0.	Deducts each category's loss: $5,000 loss, AGI $5,000. Reversed signs give $10,000 loss and AGI -$5,000. IRC 461(l).
Same +$10,000/-$5,000 on joint head, spouse no income	Loss shares [0, 0], federal person AGIs [$10,000, $0].	Allowed loss $5,000, but ownership weights still use net SE; equal fallback gives loss shares [$2,500, $2,500] and AGIs [$7,500, -$2,500]. Joint AGI $5,000 is correct. Open question for separate review: should known category-owned losses retain their owner in the federal allocator? This reconcile leaves main's allocation.
2025 Kentucky joint head ordinary SE -$5,000/SSTB +$5,000, spouse wages $40,000	Loss $0; Kentucky AGIs [$5,000, $40,000].	Federal AGIs [$2,500, $37,500]; Kentucky reverses equal federal allocation, then deducts head's $5,000: [$2,500+$2,500-$5,000, $37,500+$2,500] = [$0, $40,000]. This integration landed in #9890. Form 740 instructions p. 13.
2025 joint ages 45, wages [$1,000, $6,000], directly supplied actual IRAs [$6,000, $6,000]	Deductions [$6,000, $6,000], AGIs [-$5,000, $0]. Single wages $20,000/actual IRA $8,000 gives AGI $12,000.	Lower-earner compensation is $1,000+max($6,000-$6,000,0)=$1,000: deductions [$1,000, $6,000], AGIs [$0, $0]. Removing spouse IRA frees compensation, so head deducts $6,000. Single IRA is capped at $7,000, giving AGI $13,000. These deduction limits were already on main. IRC 219.
2025 joint ages 45, directly supplied IRS wages [$130,000, $6,000], head IRA $7,000 and traditional 401(k) $1,000	Entire IRA deducted; AGIs [$123,000, $6,000], joint $129,000.	MAGI $136,000; covered-spouse phaseout $126,000-$146,000 reduces deduction by $7,000 x $10,000/$20,000=$3,500. AGIs [$126,500, $6,000], joint $132,500. Existing main behavior. Pub. 590-A Table 1-2.
2024 Montana single age 70, wages $100,000, medical $10,000, Medicare enrollment false	Zero individual Montana AGI floor gives medical deduction $10,000.	Federal medical deduction $10,000 - 7.5% x $100,000 = $2,500; Montana's $5,500 age subtraction does not reduce the federal floor. Existing main fix; pre-2024 joint medical and joint standard formulas agree. Form 2 instructions p. 15.
2023 Montana separate spouse, supplied medical_expense_health_insurance_premiums $6,000, own self_employed_health_insurance_ald_person $1,000, other medical $500, mt_agi_indiv $40,000	Medical deduction $6,000+$500-$3,000=$3,500.	Removes previously deducted premiums: $6,000-$1,000+$500-$3,000=$2,500. Existing main fix. Form 2 instructions p. 36.
2025 West Virginia age 66, wages $4,000, taxable state refund $2,000	Senior deduction $6,000; total subtractions $8,000; WV AGI -$2,000.	Removes same-person refund before senior cap: senior deduction $4,000, subtractions $6,000, WV AGI $0. Main similarly excludes own military income and 529 contributions; at supplied AGI $6,000, military income $2,000 gives old/main senior $6,000/$4,000, and 529 $1,000 gives $6,000/$5,000. Existing main fix. IT-140 Schedule M p. 6.

Test retention and changes

No distinct test case was dropped and no YAML expected output changed. A normalized audit of all 22 original YAML files found 129 matching cases with identical inputs, outputs, periods and tolerances. Every ALD/state case from the old PR remains on main, including all old Montana medical (3), Montana standard (8), and WV (5) cases. Main's additional HSA ratio, IRA compensation, medical and WV cases remain. The HSA fallback case has main's renamed title with identical inputs/results. The merge briefly duplicated Ohio's identically named spouse-IRA case; the redundant copy was removed, leaving main's existing case intact.

The first determinism run passed 8 cases but its hash-seed child failed because main's student-loan MAGI formula reads total_unemployment_compensation, while the slim fixture captured only unemployment_compensation. The fixture now captures the total as well. It still compares the same five production formulas and float32 holders with its full-country oracle across all six hash seeds; all nine tests and expectations are retained. The repaired file passed all 9 cases, including all six hash seeds (local wall 110.52 s, pytest 74.58 s, peak reported process RSS 1,047,199,744 bytes; the failed pre-repair run was 62.45 s, pytest 6.01 s, peak 1,056,260,096 bytes). This is a test-harness repair, with no policy formula or full-model construction added. After #10221 landed the same repair on main, the second merge of main (121d534b6b) took main's version of the file, so it no longer appears in this PR's diff.

The Medicaid invariant now excludes traditional_ira_deduction alongside student-loan deductions from the head's independence comparison, because IRA deductions respond to shared compensation and joint MAGI. For example, in the $1,000/$6,000 wage household above, removing the spouse's $6,000 IRA legitimately raises the head's deductible IRA from $1,000 to $6,000. All generated households still receive the accounting, signed-netting, floor, swap and fixed $7,500 SE-reduction checks; no simulation construction or test case was added.

Fresh verification runs each file separately in the foreground, sequentially, with all five thread caps set to 4 and OMP_WAIT_POLICY=PASSIVE, using supplied Python 3.13.9 / Core 3.33.2 / pytest 9.1.1 / Hypothesis 6.168.5. No partner contract tests were edited. No microsimulation or full suite was run.

89 passed across 10 files. make format (ruff format and check) passed with no formatting changes.

File	Passed
medicaid_adjusted_gross_income_person.yaml	10
test_medicaid_household_income_invariants.py	5
medicaid_household_income.yaml	13
medicaid_magi_person.yaml	2
medicaid_person_is_required_to_file.yaml	8
test_spouse_above_the_line_deductions_invariants.py	7
test_ald_determinism.py	9
mt_medical_expense_deduction_joint.yaml	8
mt_standard_deduction_joint.yaml	9
wv_senior_citizen_disability_deduction_person.yaml	18

The new Medicaid Python file tests vectorized accounting, metamorphic relations and independent linked membership across simulation API calls; policy examples remain in variable-named YAML. It runs in the existing Rest Python shard 2/core group. The reconcile adds no test cases or simulation constructions to that file. Local macOS measurement including startup: 102.25 s wall, 36.79 s pytest, with 13.66 s total test calls. Before-reconcile timing was not measured in this checkout. Its initial BSD resource wrapper could not obtain peak memory because the sandbox denied sysctl; test execution itself passed. Other file measurements use per-process getrusage in untracked evidence. These local results do not establish Linux CI group memory usage. The repository's historical pre-isolation core group report was 13.3 GB / 11:51; current group fit remains for CI, with no added runners or parallel heavy processes.

Law and approved methodology

42 CFR 435.603(d), (e), (f) defines the included members and summation of MAGI-based income; IRC 36B(d)(2) supplies the MAGI methodology. 2025 Form 8962 instructions, line 3, p. 8 combines signed taxpayer and dependent MAGIs before flooring only the total.

Methodology ruling (Max, d1120, 2026-10-09): approved as built. A non-filer Medicaid household sums its members' signed MAGI and floors only the household total at zero (42 CFR 435.603(d)(1); Form 8962 line 3). Unlinked relatives' income counts only when positive.

axiom: TheAxiomFoundation/rulespec-us#1507 queued (42 CFR 435.603(d)(1): us/policies/medicaid/magi_household_income_pipeline.yaml already sums members' MAGI with no per-member floor; the queued companion tests pin a spouse's negative MAGI offsetting the household)

Hub microsimulation impact (current head)

A real microsimulation on the default dataset, main at the merge base 800e5bf53c against this PR's head 987402badf, for 2025 and 2026. The runs went one at a time under the shared heavy-job lock. Sums are weighted.

Output	2025 change	2026 change	2025 records changed (weighted)	2026 records changed (weighted)
medicaid	+570	+1,730	1,768 (2,569,887)	5,272 (18,083,468)
is_medicaid_eligible	+60,642	+194,970	68 (60,642)	73 (194,970)
is_chip_eligible	-19,311	-19,449	14 (19,311)	11 (19,449)
aca_ptc	-48.81M	-342.79M	8 (9,656)	10 (23,336)
household_health_benefits	+0	+0	0 (0)	0 (0)
household_net_income_including_health_benefits	-43.87M	+18.81M	1,015 (1,725,096)	2,574 (8,900,110)
state_income_tax	+0	+0	0 (0)	0 (0)
income_tax	+0	+0	0 (0)	0 (0)
household_net_income	+31,594	+5.81M	7 (44)	8 (5,668)
household_benefits	+31,594	+5.81M	7 (44)	8 (5,668)

Federal and state income tax don't change: the per-spouse deductions are main's since #9890, so this PR's remaining effect is Medicaid's signed netting of members' MAGI. Medicaid eligibility rises by 60.6k people in 2025 (68 records) and 195.0k in 2026 (73 records); no one loses it. is_chip_eligible changes for 14 and 11 records, a net fall of 19.3k and 19.4k. Total medicaid moves by only $570 and $1,730 out of about $880bn, because the model holds each state's Medicaid spending fixed: medicaid_cost_if_enrolled divides the state's calibrated spending by medicaid_slcsp_state_denominator, which over a dataset sums weight × cost index × enrollment by state. More enrollees therefore lower the per-person value for others: $398.5M moves between records in 2025 (1,709 down, 59 up) and $2,071.7M in 2026 (5,210 down, 62 up). aca_ptc falls by $48.8M in 2025 (8 records) and $342.8M in 2026 (10 records), all decreases. Net income including health benefits changes by −$43.9M and +$18.8M; net income without health benefits rises by $0.03M and $5.8M (7 and 8 records).

Comparison files are in fixes/9860-impact/ on the hub host.

[MaxGhenis]

Context for the hub from the original session, which is handing this PR over and will not push to it.

• The Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406 resolution here matches mine. Before the hub adopted the PR, I resolved the same merge locally and never pushed it. My medicaid_household_income.py formula is identical, line for line, to the one at 6c5f5cd; only one comment differs. It is Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406's parent-link membership (medicaid_non_filer_member_sum, medicaid_filer_spouse_sum, medicaid_tax_dependent_spouse_sum, the head's separate-spouse propagation) fed this PR's signed member income, with one floor on the household's total.
• Test evidence for that resolution, run 10/07 on Give each spouse their own above-the-line deductions in per-person AGI #9890 eaabd72 merged with main 07227ea (includes Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406 and Cap the educator expense deduction per eligible educator (26 U.S.C. 62(a)(2)(D)) #9833), under heavy_run.sh:
  • seam YAML (educator, loss_ald_person, own deductions, DC, parent_ids.yaml): 54 passed;
  • gov/hhs/medicaid: 5 batches passed; gov/hhs/chip and states/ca/dhcs/medi_cal: passed;
  • property tests: test_parent_links.py, test_parent_links_properties.py, test_medicaid_household_income_invariants.py, the dependent, educator and determinism files: passed.
  • The only failure was Give each spouse their own above-the-line deductions in per-person AGI #9890's spouse property test, all 7 tests. Main's Subtract Mississippi's HSA deduction once, from the account holder #9865 added health_savings_account_ald_person, which Give each spouse their own above-the-line deductions in per-person AGI #9890 then still listed as equally divided. The hub's Give each spouse their own above-the-line deductions in per-person AGI #9890 head handles that now.
• The impact tables in this body (2024/2026) predate Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406. They compare Give each spouse their own above-the-line deductions in per-person AGI #9890 36058ba with this branch at d4b561e. Medicaid MAGI household composition from co-resident parent ids (fixes #9404) #9406 changes Medicaid household membership, so they need a rerun at the current heads.

🤖 Generated with Claude Code

[MaxGhenis]

US + core hub merge audit, us#9860 at 987402badf5970094434183ea1424c425b06d436 (squash): Net household members' Medicaid MAGI before the zero floor, and give each spouse their own deductions.

• Review: GPT-6.1 Sol r5 APPROVE at 987402b (doc-only delta; reviews-gen/us-9860/review-r5.md), after r3 full review APPROVE of the reconciled PR at 121d534 (no P1/P2, METHODOLOGY no) and r4 at f5c83ea. Methodology approved by Max as built (d1120).
• 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 800e5bf53c against this head, in the description): 2025: medicaid +570 (1,768 records); is_medicaid_eligible +60,642 (68 records); is_chip_eligible -19,311 (14 records); aca_ptc -48.8M (8 records); household_net_income_including_health_benefits -43.9M (1,015 records); household_net_income +31,594 (7 records) | 2026: medicaid +1,730 (5,272 records); is_medicaid_eligible +194,970 (73 records); is_chip_eligible -19,449 (11 records); aca_ptc -342.8M (10 records); household_net_income_including_health_benefits +18.8M (2,574 records); household_net_income +5.8M (8 records).
• Axiom: TheAxiomFoundation/rulespec-us#1507 queued (42 CFR 435.603(d)(1): us/policies/medicaid/magi_household_income_pipeline.yaml already sums members' MAGI with no per-member floor; the queued companion tests pin a spouse's negative MAGI offsetting the household)