Fix Arkansas 2025 and 2024 income tax minus-adjustment schedules against the DFA cards

[MaxGhenis]

Summary

• 2025: rates/main/reduction.yaml jumped from the $26,400 row ($420) straight to $95,001. It skipped the three DFA rows for $94,701–$95,000, so taxable income in that band got the $420 reduction instead of $399.30, $389.30 or $379.30 and was undertaxed by $20.70–$40.70. The 2025 column also used whole-dollar amounts instead of DFA's published cents: 112, 224, 288, 420, then 369, 359, … 89 instead of 111.98, 223.97, 287.97, 419.96, then 369.30, 359.30, … 89.30. This PR re-keys the whole 2025 column to the 2025 Indexed Tax Brackets card.
• 2024 (found while verifying the other years): the card's asterisked row "*$94,301 *$94,500 3.90% $197.40" is intentionally $200 wide. The 2024 column started the next row at $94,401, so every income from $94,401 to $95,500 took the next $10 step down and was overtaxed by $10. This is the same defect AR 2023 bracket-adjustment column mistranscribed: missing $89,901 row, un-encoded asterisked $200 ranges, duplicate 91,701/91,801 thresholds (40/44 rows off) #9026 fixed for 2023.
• Moves the 2025 booklet reference in reduction.yaml and rate.yaml from #page=26 to #page=30. #page=26 matched the printed page number, but PDF page 26 is "For Taxpayers' Information". The Regular Income Tax Table starts at PDF page 30. Both files now also cite the 2025 TaxBrackets card.
• Tests added or updated:
  • 17 YAML tests covering 2025's $94,701–$95,000 band and its boundaries, the 2025 ladder ends, and the 2024 asterisked row.
  • 4 existing 2025 tests move from rounded amounts to DFA cents, with the margin tightened from 1 to 0.01. The joint test file gets the same cents correction in its comments and in one expected value.
  • A new pytest, policyengine_us/tests/test_ar_income_tax_schedule.py; see "Invariants" below.

Before and after

ar_income_tax_before_non_refundable_credits_indiv, computed as 3.9% × taxable income − the row's minus adjustment:

Year	Taxable income	DFA row	upstream/main	This PR
2025	94,700	$26,400–$94,700, $419.96	3,273.30	3,273.34
2025	94,751	$94,701–$94,800, $399.30	3,275.29	3,295.99 (table: $3,296)
2025	94,851	$94,801–$94,900, $389.30	3,279.19	3,309.89 (table: $3,310)
2025	94,951	$94,901–$95,000, $379.30	3,283.09	3,323.79 (table: $3,324)
2025	95,001	$95,001–$95,100, $369.30	3,336.04	3,335.74
2024	94,450	$94,301–$94,500, $197.40	3,496.15	3,486.15 (table: $3,486)
2024	95,500	$95,401–$95,500, $97.40	3,637.10	3,627.10

"Table" is DFA's Regular Income Tax Table value for the $100 band. DFA computes it at the band's midpoint and rounds to the dollar: 2025, 2024.

Why DFA's cents, not whole dollars

The rounded 2025 amounts weren't a documented choice. #7362 cited the booklet's whole-dollar tax table, not the TaxBrackets card. The 2021–2024 columns all carry DFA's cents. The card says "If you use a formula to calculate Arkansas income tax, the results must match the table exactly", and its worked examples subtract the cents amount ($75,950 x .039 = $2,962.05 - $419.96).

The rounded amounts don't reproduce DFA's table either. I parsed all 949 bands of the 2025 Regular Income Tax Table (PDF pages 30–34) and evaluated the model at each band's midpoint, rounding half up:

2025 Regular Income Tax Table	upstream/main	This PR
Bands missed because of the three skipped rows	3	0
Bands missed by $1 because of rounding (between $75,501 and $99,501)	35	0
Total missed of 949	38	0

The same check on 2024: upstream/main misses the 11 bands from $94,401 to $95,501; this PR misses none.

Rows are shared across years

After the fix, 2025 rows 5–36 hold $94,701 + $100k with amount $399.30 − $10k. They line up exactly with the 2026 rows from Act 1 of 2026 ($94,701 + $100k, amount $369.90 − $10k). The two rows that became finite in 2025 ($97,701 and $97,801) already had explicit 2026-01-01: .inf, so 2026 inherits nothing new.

Verification

• Every integer taxable income, 0–110,000 (plus 150,000 and 1,000,000), compared with the source. 2021–2025 use the DFA Indexed Tax Brackets cards. 2026 uses Act 1 of 2026 § 1 (A.C.A. § 26-51-201(a)(4)(A)–(C)):

  Year	2021	2022	2023	2024	2025	2026
  Source rows	12	52	50	37	37	Act
  Mismatches, upstream/main	0	0	0	1,100	104,403	0
  Mismatches, this PR	0	0	0	0	0	0

  Independent agents re-parsed each year's source with pdfplumber and pdftotext. They found 0 mismatches on this branch for 2021–2025 and the same upstream/main mismatch ranges.

• Loaded scales, 2015–2035 (Jan 1 and Jul 1 of each year): only the 2024 and 2025 reduction scales change. rate.yaml's loaded values are identical; only its references changed. rate.yaml needs no new rows, because the rate is a flat 3.9% from $26,400 up in both years.

• Tests:

  • policyengine-core test policyengine_us/tests/policy/baseline/gov/states/ar → 513 passed.
  • The new pytest → 18 passed.
  • Against upstream/main's parameters:
    • 20 of the 46 cases in the indiv YAML file fail: the 4 re-centred 2025 cases, the 12 new 2025 cases, and 4 of the 5 new 2024 cases. The 2024 $95,501 boundary case passes on both, because that row was already correct.
    • The pytest's 2024 and 2025 differential cases fail.

• Open PRs nearby:

  • Fix duplicate 2024 Arkansas low-income table threshold and guard scale thresholds #9599 edits the low-income tables in the same folder and touches none of these files. Its new duplicate-threshold guard, run against this branch, reports nothing from rates/main. All 25 of its findings are the rows Fix duplicate 2024 Arkansas low-income table threshold and guard scale thresholds #9599 itself fixes.
  • Fix Arkansas 2014-2020 income tax rate schedules #9611 (2014–2020 schedules) also edits reduction.yaml and rate.yaml. git merge-tree of the two PRs has no conflicts, and the merged tree still has 0 mismatches for 2021–2026.

• Citation check: arkleg's status page says "Notification that HB1001 is now Act 1" (5/6/2026), so the existing Act 1 citation is correct.

Invariants

policyengine_us/tests/test_ar_income_tax_schedule.py checks these for every integer taxable income from 0 to 150,000 (plus 250,000 and 1,000,000) in 2021–2026:

1. Differential: the model's rate × TI − reduction equals an independent, compact transcription of each DFA card or Act 1 of 2026. The transcription is the base rows plus a $100/$10 ladder with explicit asterisked $200 rows. The check allows at most half a cent. It catches this whole class of bug: skipped rows, ignored asterisked rows, shifted rows.
2. Tax is non-negative.
3. Tax never falls as income rises, with one exception labeled as intended. Act 1 of 2026 itself sets $3,136.70 at $94,700 under (a)(4)(A) and $3,134.04 at $94,701 under (a)(4)(B) less the $290 (a)(4)(C) adjustment, a $2.66 drop written into the statute. The test pins that drop exactly.
4. Row thresholds strictly increase: in each year, the reduction.yaml rows carry strictly increasing finite thresholds, with every .inf row after them. The test reads each row's own threshold. The loaded scale can't reveal a mis-keyed row, because add_bracket sorts thresholds and sums duplicate amounts. This check flags the pre-AR 2023 bracket-adjustment column mistranscribed: missing $89,901 row, un-encoded asterisked $200 ranges, duplicate 91,701/91,801 thresholds (40/44 rows off) #9026 2023 duplicates (91,701 and 91,801), which a loaded-scale check passes. The independent review caught that the first version of this test checked the loaded scale.

Not changed

• 2026 lower brackets (112 / 224 / 288 / 367.20): Act 1 of 2026 gives marginal rates only, and DFA has not published a 2026 TaxBrackets card. The 2026 amounts follow the Act's "From" bounds. DFA's 2021–2025 cards instead count from the "Less Than or Equal To" bound (111.98 = 2% × 5,599). If DFA's 2026 card keeps that convention, the lower 2026 amounts would become 111.98 / 223.97 / 287.97 / 367.16, a 2–4 cent change. That is left for when the card is published.
• The booklet's shortcut above $100,000 ("$3,809 + 3.9% of the excess over $100,000" for 2025; "$3,811 + …" for 2024) is $1.70 and $1.60 below the card's bracket formula (0.039 × TI − 89.30 / − 87.40). PolicyEngine follows the bracket formula, as it did before.

axiom: TheAxiomFoundation/rulespec-us#1410 debt. Axiom has Arkansas income tax only for TY2026 (the Act 1 of 2026 pilot, bounded to 2026). #1405 covers TY2014–2020. #1410 records the TY2021–2025 schedule, including these 2024 and 2025 rows.

🤖 Generated with Claude Code

[MaxGhenis]

Admin-merged (squash, --match-head-commit 3045506) under Max's 2026-09-26 ruling (d399: obvious improvements need no permission; admin-merge when the only block is a required review only Max could give). Gates: gh pr checks exit 0 (37 pass), MERGEABLE, no CHANGES_REQUESTED. Approvals: an independent adversarial Opus 5.5 review of the final head (workflow wf_b23156c4), which re-checked the 2025/2024 values against the DFA Indexed Tax Brackets cards and found no blocker or major findings, plus the session-management main session's own diff review. Closes decision d396.