[MaxGhenis]

Part of #10442. That issue also asks for California Case 19 to derive the federal credit. The case arrives with #10305, so the issue stays open until #10305 lands (see "California Case 19" below).

The bug

gov.aca.required_contribution_percentage.initial and .final stopped at 2026 (Rev. Proc. 2025-25), so 2027 reused the 2026 table. At 200% FPL the model used 6.60% where the law for 2027 says 6.78%. Every 2027 percentage is higher than its 2026 counterpart, so the model set the required contribution too low for every household with positive income at or below 400% FPL. For those households it overstated the 2027 federal premium tax credit wherever that credit was above $0 under the 2026 table. Where the benchmark premium is below the required contribution under both tables, the credit is $0 either way. Programs built on the federal credit inherit the error. The California Premium Subsidy, for example, is max(0, slcsp − aca_ptc − ca_pct × income), so while that subsidy is above $0, each dollar of extra federal credit is a dollar less of it (#10305).

The fix

Rev. Proc. 2026-26 §3.01 sets the table for taxable years beginning in 2027. I checked each value against the PDF:

Household income (% FPL)	2026 (Rev. Proc. 2025-25)	2027 (Rev. Proc. 2026-26)
Less than 133%	2.10%	2.15%
At least 133% but less than 150%	3.14% → 4.19%	3.23% → 4.3%
At least 150% but less than 200%	4.19% → 6.60%	4.3% → 6.78%
At least 200% but less than 250%	6.60% → 8.44%	6.78% → 8.66%
At least 250% but less than 300%	8.44% → 9.96%	8.66% → 10.22%
At least 300% but not more than 400%	9.96%	10.22%

• initial.yaml and final.yaml get 2027-01-01 rows, and all three table files cite Rev. Proc. 2026-26 (file page 3).
• threshold.yaml needs no new values. The 2027 bands are the same six as 2026: 133, 150, 200, 250, 300 and 400 percent.
• The Rev. Proc. 2025-25 hrefs gain #page=3, where the 2026 table is, per the reference rules.
• §3.02 also sets the 2027 Required Contribution Percentage (employer-coverage affordability, §36B(c)(2)(C)) at 10.22%. PolicyEngine-US has no parameter for it: affordability is the input offered_aca_disqualifying_esi. A search of parameters/ and variables/ for 9.96% (the 2026 Required Contribution Percentage), "affordab", 36B(c)(2)(C) and Rev. Proc. 2025-25 found no other 2026-only ACA indexing parameter.

Tests

YAML, with expected values from the Rev. Proc., not from model output.

• aca_required_contribution_percentage.yaml: 15 cases in 2027. Cases at 100% and 132% are flat 2.15%. Each edge from 133 to 300 percent gives that band's initial percentage, and 400 percent gives 10.22%. Interior points sit at 140, 175, 199, 225, 275 and 350 percent, and 401 percent gives 0.
• aca_ptc.yaml: 9 cases in 2027 that derive the percentage, at 100, 133, 150, 200, 275, 350 and 390 percent (the last floors at $0). Two cases derive the FPL fraction from income and the 2026 HHS guideline: one person at $31,920 (200%) and a couple at $54,100 (250%). The 200% case is the Covered California OE27 example: $8,256 − 6.78% × $31,920 = $6,091.82.
• Against main's parameters, 14 of the 15 new percentage cases and 8 of the 9 new credit cases fail. The two that pass on main pass for the right reason: 401% is $0 under both tables, and the 390% credit floors at $0 under both.

Existing 2027 expectations. None move. I listed every YAML case whose period is 2027 or later (305 files) and kept those whose outputs depend on aca_required_contribution_percentage. To find those, I built a static reverse-dependency closure over formulas, adds and subtracts, and the list parameters passed to add(). That left 17 files with 35 cases. Each file passes unchanged on this branch, one file per process: those cases supply aca_ptc as an input or test eligibility and program years. No partner contract test has a 2027 or later case.

California Case 19. #10305 adds amount Case 19 to ca_premium_subsidy.yaml, the OE27 example, and supplies aca_ptc: 6_091.824 because main lacks this table. This PR lands before #10305, so #10305, landing second, changes Case 19 to derive aca_ptc. The expected values are a federal credit of $6,091.82 and ca_premium_subsidy of $478.80. #10442 stays open until then.

Invariants

These hold for every income, in every year from 2018 through 2027, against that year's table. policyengine_us/tests/core/test_aca_required_contribution_percentage_properties.py checks them with Hypothesis:

1. At each band's lower edge, the applicable percentage equals that band's initial percentage. At exactly 400 percent it equals the top band's final percentage.
2. Within a band the percentage is the straight line from the band's initial to its final percentage, rounded to the nearest hundredth of a percent (26 CFR 1.36B-3(g)(1)). So it stays between the band's initial and final percentages, reaches the final percentage just below the next edge, and at every income is within half a rounding step (0.005 percentage point) of the line. The bound holds whether or not the model rounds (it doesn't yet; see below).
3. The percentage never decreases as income rises from 0 to 400 percent, including across edges. Before rounding, the line is continuous within each band and the only jump between bands is at 133 percent: 2.15% → 3.23% in 2027. Rounding adds steps of 0.01 percentage point.
4. The required contribution (income × percentage) never decreases with income. So with a fixed benchmark premium, the credit never rises with income up to 400 percent. This is checked for 2026 and 2027, from income in dollars through the FPL fraction. Every example includes four fixed incomes ($10,000, $25,000, $40,000 and $60,000, one person) in four different bands, so no example compares only equal incomes.

The test also asserts that every year's band edges are among the edges it probes, so a mistyped edge can't pass by being skipped.

Mutation checks (each mutant run against the property test at 83c7d8c, one process at a time, RSS-capped at 2 GB):

• Caught: a 2027 final percentage below its initial (property 2), 2027 rows out of order (property 3), a mistyped 2026 edge, 1.35 for 1.33 (the edge guard), an interpolation that overshoots by 20% (property 2), and a bow of up to 0.008 percentage point in the sloped bands that keeps every endpoint and the order (property 2's line bound; Sol's second review found that the earlier pairwise bound let it through).
• Passes, as intended: rounding the percentage to the nearest hundredth of a percent, as 26 CFR 1.36B-3(g)(1) requires. Sol's first review found that the first version of the test rejected it.
• Survived: searching edges with side="left". It is an equivalent mutant: the model stores the FPL fraction in float32, and float32(1.33) is just above 1.33, so 133% still lands in the upper band. At every other edge the two bands meet at the same value.

Cost (Python test, measured twice locally on a shared macOS host, not Linux CI): 4–6 s of test calls, 7–12 s in pytest and 29–36 s wall time with the model import; peak memory footprint 1.7 GB. It runs in the core group of the rest Python tests, like test_aca_ptc_filing_gate_invariants.py.

Impact (2027)

The US hub ran the full default dataset for 2027: main at the merge base 41d218c474 against this PR's head 83c7d8cd1a, one run at a time under the shared heavy-job lock. Sums are weighted.

• The premium tax credit (aca_ptc, per tax unit) falls by $2.00bn across 8,390 records (16.3 million weighted tax units); none rises, and the largest single drop is $548.
• The credit tax units take up (used_aca_ptc) falls by $805M, and their net marketplace premiums rise by the same amount.
• State subsidies that top up the federal credit rise by $105M in all, mostly California, Maryland, New Mexico and Massachusetts. The Basic Health Program falls by $35M.
• Income tax, state income tax and household_net_income don't change. SPM unit net income falls by $805M, and 10 records (0.1 weighted people) move into poverty.

Output	2027 change	2027 records changed (weighted)
aca_ptc	-2.00bn	8,390 (16,268,136)
assigned_aca_ptc	-831.14M	2,745 (5,668,900)
used_aca_ptc	-805.31M	2,718 (5,581,312)
marketplace_net_premium	+805.31M	2,718 (5,581,312)
basic_health_program	-34.57M	336 (692,558)
ca_premium_subsidy	+22.86M	119 (421,410)
co_premium_assistance	+0	2 (0)
md_premium_assistance	+35.77M	134 (252,215)
nm_premium_assistance	+16.40M	182 (170,993)
vt_premium_assistance	+0	0 (0)
ct_covered_connecticut	+629,568	18 (18,605)
ma_connector_care	+29.67M	177 (265,266)
nj_njhps	+0	0 (0)
wa_cascade_care_savings	+21,139	2 (1,493)
healthcare_benefit_value	-760.37M	3,097 (6,234,798)
household_net_income_including_health_benefits	-760.37M	3,097 (6,234,798)
spm_unit_net_income	-805.31M	2,591 (5,414,091)
person_in_poverty	+0	10 (0)
state_income_tax	+0	0 (0)
income_tax	+0	0 (0)
household_net_income	+0	0 (0)
household_benefits	+0	0 (0)

Comparison files are in fixes/10443-impact/ on the hub host.

This is a model-to-law fix, so the size of the aggregate change doesn't need approval.

Not in this PR

• 26 CFR 1.36B-3(g)(1) rounds the interpolated applicable percentage "to the nearest one-hundredth of one percent". The model doesn't round, in any year: at 140% FPL in 2027 it uses 3.6706% where the regulation gives 3.67%. This PR's interior YAML points are chosen so the expected value is the regulation's rounded figure and the unrounded value is within 0.001 percentage points of it. I'll propose the rounding as its own change.

axiom: us/guidance/irs/rev-proc-2026-26 TheAxiomFoundation/rulespec-us#1909 queued (the corpus scope holding Rev. Proc. 2026-26 is not in rulespec-us's pinned release)

🤖 Generated with Claude Code

[MaxGhenis]

Merged at head 83c7d8cd1a97604afc65ea4f2e634158f82be1e2 with --admin --match-head-commit. The only open block was a required approving review that no one but the repo owner could give. This is a model-to-law fix, so no owner decision was needed. The gates:

• Independent review: GPT-6.1 Sol, three rounds.
  • Round 1 (05310b2) requested changes: the property test rejected the 26 CFR 1.36B-3(g)(1) rounding. Fixed in aa7c814.
  • Round 2 (aa7c814) approved with two nits. Both were fixed in 83c7d8c with a pointwise interpolation-line bound and continuity wording.
  • Round 3 approved this head. Its one nit, a mutant that also bowed flat bands, was rerun with the bow limited to sloped bands. It fails at the line-bound assertion.
• CI: gh pr checks exits 0, and the latest "Pull request" run at this head concluded success. MERGEABLE, not a draft, no CHANGES_REQUESTED.
• Impact: the hub's full-file 2027 run (main 41d218c vs this head), in the body. aca_ptc falls $2.00bn, used_aca_ptc falls $805M, and state top-ups rise $105M. Income taxes and household_net_income are unchanged.
• Partner tests: untouched, and none has a 2027 or later case.
• Follow-up owed: Decide CA Premium Subsidy eligibility on the exact FPL ratio and adopt the Final 2027 design #10305 lands second, so it changes California Case 19 to derive aca_ptc ($6,091.82 federal, ca_premium_subsidy $478.80). Add the 2027 premium tax credit applicable percentages (Rev. Proc. 2026-26) #10442 stays open until then.
• Separate follow-up: rounding the applicable percentage to the nearest hundredth of a percent (26 CFR 1.36B-3(g)(1)), which the model has never done.

🤖 Generated with Claude Code