[MaxGhenis]

Draft. Done: the fix and YAML cases. Pending: full-file impact, Axiom parity line, GPT-6.1 Sol review.

Bug

vt_premium_assistance_eligible compared the truncated aca_magi_fraction (np.floor(100 * magi / fpg) / 100) with <= 3.0. Truncation maps every income from 300.00% to 300.99% of the poverty line to 3.00, so all of them passed. 33 V.S.A. § 1812(a)(1) covers "an individual or family eligible for federal premium tax credits under 26 U.S.C. § 36B with income less than or equal to 300 percent of federal poverty level". No source I found prescribes truncating the percentage for this test. (A research pass reported the same inclusive wording in HBEE Part 7 § 56.03; the agency PDFs now return 404, so that quote is not relied on here.)

Executed on main (645fcd3; the file is unchanged at 740e1eb), single filer in Vermont, 2026, 2025 one-person guideline $15,650, is_aca_ptc_eligible supplied:

Income	Exact % of FPL	aca_magi_fraction	Eligible on main	Eligible here
$46,950	300.000%	3.00	yes	yes
$46,950.47	300.003%	3.00	yes	no
$47,000	300.319%	3.00	yes	no
$47,105.56	300.994%	3.00	yes	no
$47,106.50	301.000%	3.01	no	no

The federal credit ends at 400%, so these households are federally eligible and the Vermont gate alone decided them.

Fix

Decide eligibility on the exact ratio max(aca_magi, 0) / tax_unit_fpg(period.last_year), the pattern of #10305. vt_premium_assistance and its required-contribution lookup still use aca_magi_fraction; only the eligibility edge moves.

Invariants

For every input, with the program in effect and an APTC-eligible member:

1. Band = statute. Eligible if and only if r <= 3, r = max(MAGI, 0) / FPG(prior year).
2. Only the sliver changes. Relative to main, eligibility can only be lost, and only for r in (3.00, 3.01). Deduced: floor(100r)/100 <= 3 holds exactly when r < 3.01.

No Hypothesis test is added: the gate is one comparison on a ratio, and the YAML cases below hit the limit itself, 0.003 and 0.994 points above it, and 300.32% end to end. The repo's test-design rules keep such cases in YAML.

Tests

• vt_premium_assistance_eligible.yaml: cases that supplied aca_magi_fraction now set aca_magi and a prior-year tax_unit_fpg of $10,000 (the gate no longer reads the fraction). Case 10, which supplied 3.0001, a value truncation can never produce, now supplies $30,001 against $10,000 and exercises the fix. New income-derived cases with the real guideline: $46,949 (299.994%) and $46,950 (300%) eligible, $46,950.47 (300.003%) and $47,105.56 (300.994%) ineligible.
• integration.yaml Scenario 2 expected a $47,000 single filer (300.32%) to stay eligible; its comment said "47_000 sits just above it but the model still rounds aca_magi_fraction to 3.00, so the inclusive ceiling holds". It now uses $46,950 (exactly 300%; VPA $704.25), and Scenario 2b keeps $47,000 as an ineligible case (VPA $0, federal APTC unchanged at $10,906.80).
• integration.yaml Scenario 8 and vt_premium_assistance.yaml Case 13 supplied a 3.00 fraction with $100,000 of MAGI for three people, which is 375% of the $26,650 guideline. They now use $79,950, exactly 300% (VPA $1,199.25).
• Mutation check: main's gate fails exactly Scenario 2b and eligibility Cases 10, 12 and 13.
• Local differential sweep (a script, not committed, per the repo's test-design rules): 256 tax units per year, 60% with income in cents within $2 of an edge and the rest uniform, guidelines $10,000-$80,000 including $20 multiples, comparing the eligibility gate with an exact-rational reading of the source on the float32 values the model receives. 0 mismatches here; on main the same draws give 63 mismatches, all in the truncation sliver.
• Local run: the three premium-assistance YAML files, 39 passed.

Impact

Pending: full-file microsimulation via the US hub impact queue, 2026 and 2027.

Coordination

#10041 (draft) changes aca_magi_fraction's denominator to a new aca_fpg; if it lands, this gate's exact ratio should use the same denominator. Part of the aca_magi_fraction audit (table: #10329) begun with #10305; federal edge in #10312.

axiom: pending

🤖 Generated with Claude Code