[MaxGhenis]

The 2021 fully refundable CTC omitted the first ARPA reduction and could refund credit for other dependents (ODC). This corrects the Schedule 8812 split, including a solar-credit ordering case that previously overstated the federal refund by $500.

Using child-only credit C, ODC O, the capped first ARPA reduction F, and ordinary phaseout P:

• Total before the tax limit: T = max(C − F + O − P, 0) (line 12).
• Refundable: R = max(C − F − P, 0) (line 14b).
• Nonrefundable: N = min(O, T) (line 14a), retained before the §26(a) tax limit applied on line 14d.

The first reduction retains its caps at the enhancement above $2,000 per qualifying child and 5% of the ordinary-minus-ARPA threshold difference. The partial-refundability branch and existing refundable-maximum guard are preserved. The ordinary phaseout threshold for surviving spouses is also corrected to $200,000 from 2018 onward; their separate 2021 ARPA threshold remains $150,000.

Sources: 2021 IRC §24(h) and (i), PDF pages 57–58, 2021 Schedule 8812, page 1, and Line 5 Worksheet, instructions page 4.

Implementation plan:

• Correct the refundable split and document pre-limit nonrefundable entitlement.
• Add the three U5 regressions: ARPA plus ODC, the floor after ordinary phaseout, and solar ordering.
• Add vectorized partition, child-only refund, and other-year compatibility properties.
• Add Oklahoma binding-tax-limit coverage and surviving-spouse threshold/cap tests.
• Review downstream consumers. Iowa's 2021 add-back of non_refundable_ctc (a 2023 workaround for a negative split) is removed, so Iowa deducts the federal tax after the credit for other dependents.
• Add ARPA threshold-boundary and joint-return cases with both phaseouts and an ODC (review S2).
• Finish validation and record final counts below.
• Hub: run the impact analysis and complete parity.

Validation (after review, at 0a9022de26, main merged):

• YAML, one file per process: refundable_ctc.yaml 26, ctc.yaml 2, ctc_arpa_phase_out_cap.yaml 3, ctc_phase_out_threshold.yaml 6, ia_fedtax_deduction.yaml 5, ia_integration.yaml 9, ok_federal_ctc.yaml 6. All pass.
• Python: test_ctc_2021_refundable_split.py 7 and test_source_conflicts.py 103. Both pass. The grid is the one check that the 2021 branch leaves 2018, 2020, 2022, 2025 and 2026 unchanged across combinations of credits, reductions, caps and the section 911 bar. It takes about 2 s (1.42 s in CI at the reviewed head).
• Restoring the old split fails five of the six new boundary cases. Restoring Iowa's add-back fails the new Iowa case.
• The CI status of this head is in the hand-off comment. The full suite passed at the reviewed head cffa308032, including every file the earlier list called unfinished.

YAML files run separately through Core's official test runner in sequential foreground processes. No microsimulation was run; the hub handles impact analysis.

axiom: TheAxiomFoundation/rulespec-us#1856 queued for §24(i)(1)(B), (i)(3) and (i)(4), with surviving-spouse companion rows for the 24(h)(3) threshold (us:statutes/26/24/h#ctc_phase_out_threshold_under_subsection_h; no test yet runs a surviving spouse), blocked on ingesting P.L. 117-2 §9611 (TheAxiomFoundation/axiom-corpus#744), a re-pin to a release with the 2021 Schedule 8812, and the derived filing-status chain (TheAxiomFoundation/rulespec-us#1399).

---

Generated by Claude Code

[DTrim99]

PR Review: Fix the 2021 CTC split and surviving-spouse phaseout threshold

PR #10306 (author: MaxGhenis), head cffa308032. This is a first review. The PR makes three changes:

• The 2021 refundable split. In 2021 the refundable CTC now subtracts the first ARPA reduction as well as the ordinary phaseout. So the credit for other dependents (ODC) stays nonrefundable.
• The surviving-spouse threshold. It sets the ordinary phaseout threshold for surviving spouses to $200,000 from 2018 (it was $400,000).
• Documentation and tests. It documents non_refundable_ctc as the amount before the §26(a) limit and rewords two Iowa comments. It adds 15 YAML cases across five files, a seven-test Python grid and a changelog fragment.

The PR has no comments or reviews yet. It touches no partner test file.

Path prefixes:

• V/ = policyengine_us/variables/gov/irs/credits/ctc/
• T/ = policyengine_us/tests/policy/baseline/gov/irs/credits/ctc/

Sources read for this review (all fetched 2026-10-21):

• Statute: 26 U.S.C. 24, 2021 edition (govinfo HTML), subsections (b), (h), (i) and (j). I also read the PR's cited subtitle A PDF, file pages 57, 58 and 81.
• IRS:
  • 2021 Schedule 8812, lines 4-14i.
  • The 2021 Instructions for Schedule 8812, Line 5 Worksheet.
  • 2021 Form 5695 and its instructions.
• Iowa: 2021 IA 1040 expanded instructions, file pages 40-41 (federal tax addition and deduction).
• rulespec-us: main at 5796861e58, which has us/statutes/26/24.yaml, 24/d.yaml and 24/h.yaml. I also read issues Remove calculated variables from input/income.py #1829 and Add hide-input cell tags #1399.

Summary

What is right.

• The law.
  • First reduction. §24(i)(4)(A) reduces "the credit allowable under subsection (a) (determined without regard to subsection (b))" by "$50 for each $1,000 (or fraction thereof)" above the applicable threshold. Under (i)(4)(B) that is "$150,000, in the case of a joint return or surviving spouse (as defined in section 2(a))", $112,500 for a head of household and $75,000 otherwise.
  • Its cap. Under (i)(4)(C), the reduction may not exceed "the lesser of (I) the applicable credit increase amount, or (II) 5 percent of the applicable phaseout threshold range". The increase amount is the $3,000/$3,600 enhancement over the $2,000 credit. The range runs from the (b)/(h)(3) threshold down to the (i)(4) threshold.
  • Order. (i)(4)(D) applies subsection (b) to "the credit allowable under subsection (a) (determined after the application of subsection (i)(4)(A)". So the ordinary $400,000/$200,000 phaseout comes second.
  • Refundable part. §24(i)(1)(B) allows under subpart C "so much of the credit determined under subsection (a) ... as does not exceed the amount of such credit which would be so determined without regard to subsection (h)(4)", that is, without the ODC.
  • The form. Schedule 8812 implements this as line 12 = line 5 (child credit after the first reduction) + line 7 (ODC) − line 11 (ordinary phaseout). Then "14a Enter the smaller of line 7 or line 12" and "14b Subtract line 14a from line 12".
• The code matches the law and the form.
  • Total. ctc = max(0, base + ODC + (ARPA enhancement − F) − P), where F is the first reduction and P the ordinary phaseout. That is line 12.
  • Refundable. In 2021, refundable_ctc = min(max(0, C − F − P), ctc), where C is the sum of ARPA child amounts (ctc_refundable_maximum). That equals line 14b in every branch: if C − F − P ≥ 0, then 14a = ODC and 14b = C − F − P; otherwise 14a = line 12 and 14b = 0.
  • Nonrefundable. non_refundable_ctc = ctc − refundable_ctc is line 14a.
  • The bug fixed. The old code left out F, so it refunded up to F of the ODC.
• Surviving spouses.
  • The law. §24(h)(3): "$400,000 in the case of a joint return ($200,000 in any other case)". Before 2018, §24(b)(2)(B) gives $75,000 to "an individual who is not married", and the parameter already had that.
  • The forms. Schedule 8812 line 9 reads "Married filing jointly—$400,000 / All other filing statuses—$200,000". The Line 5 Worksheet uses "Qualifying widow(er) — $2,500" (line 6) and "Married filing jointly or Qualifying widow(er) — $150,000" (line 8).
  • The PR. It gives a surviving spouse $200,000 from 2018 and keeps the $150,000 ARPA threshold. So the 2021 cap becomes 5% × $50,000 = $2,500 instead of $12,500.
• Rounding. Both reductions use ceil(excess / 1_000) * 50 on MAGI including the §911/931/933 exclusions, as §24(b)(1) requires.
• Other years. The refundable change sits inside if ctc.refundable.fully_refundable, which is true only for 2021. The surviving-spouse change applies from 2018, as the law requires. The Python grid confirms the legacy §24(d) branch is unchanged for 2018, 2020, 2022, 2025 and 2026.
• Tests. I hand-checked 8 cases (about 35 expectations), listed under Verified correct. CI passed every changed file.
• Batch checks. On the tree merged with main, the PR passes the source-conflict ratchet (checked statically). It has no state-gated key and changes no income-source list.

There are no critical findings and three should-address items:

• S1. Invalid Axiom line. axiom: pending: hub files parity is not an allowed form. We flagged the same form on Correct Minnesota UI taxable wage bases for 2014-2026 #10319 and Fix CalEITC child counts for ITIN holders #10322. rulespec-us doesn't encode §24(i), and no open issue covers (i)(1)(B) or (i)(4).
• S2. No boundary or joint-return coverage of the new split. No new case sits on an (i)(4) threshold, where the old code first went wrong. None is a joint return, so the $150,000 and $400,000 joint thresholds are never tested together with an ODC.
• S3. Iowa's 2021 add-back of non_refundable_ctc is now an obsolete workaround. It makes Iowa's federal-tax deduction ignore the ODC. Where §26(a) limits the ODC, the PR now makes Iowa add back more than was netted. That contradicts the PR body's claim that Iowa's calculation is preserved.

There are also three suggestions: the Python test, references and comment housekeeping, and the stale PR body.

Status.

• CI: 27 checks pass and 3 are skipped (the superseded umbrella jobs). None is pending or failed. CI's base predates Tie source-conflict markers to the register shared with Axiom #10287, so I checked the ratchet statically.
• Branch: 4 commits ahead of main and 26 behind, and it merges cleanly.

Critical (must fix)

None.

Should address

S1. The Axiom line axiom: pending: hub files parity is not a valid form

• Where: the last line of the PR body.
• The rule (CLAUDE.md, "Axiom Parity"): a PR that adds, updates or fixes policy needs one of encoded-correct, <rulespec PR> encoded, <rulespec issue> queued or n/a: <reason>. pending is not one of them.
• Why n/a doesn't fit: §24(i) and §24(h)(3) are statutory. The PR changes 2021 refunds for families with other dependents above an ARPA threshold. It also changes the 2018+ CTC for surviving spouses with MAGI between $200,000 and $400,000.
• Why encoded-correct isn't available for §24(i):
  • rulespec-us main (5796861e58) has no us/statutes/26/24/i module. The 24.yaml summary covers subsection (h) and never mentions (i).
  • Remove calculated variables from input/income.py #1829 (open, pe-parity) mentions §24(i) only for (i)(1)(A), which turns off the §24(d)(3) bar. It doesn't cover (i)(1)(B), (i)(3) or (i)(4).
  • My searches for "24(i)", "2021 child tax credit", "ARPA child", "Schedule 8812" and "credit for other dependents" found nothing else.
• §24(h)(3) is already encoded. us:statutes/26/24/h#ctc_phase_out_threshold_under_subsection_h returns $400,000 when filing_status == 1 and $200,000 otherwise, from 2018-01-01. That gives a surviving spouse $200,000, provided code 1 means a joint return only. Its companion tests have no surviving-spouse case, and Add hide-input cell tags #1399 notes that filing status is not yet derived upstream.
• Severity: should. An invalid Axiom line is a merge blocker, not a model error.
• Fix: file an open, pe-parity-labelled, dispatch-ready issue in TheAxiomFoundation/rulespec-us, modelled on Add AMT parameters for 2020-2023 #1467. Then change the line to axiom: us:statutes/26/24/h#ctc_phase_out_threshold_under_subsection_h encoded-correct; TheAxiomFoundation/rulespec-us#<n> queued. Alternatively, run the signed encoder and cite the PR as encoded. The issue needs:
  • Target module: a new us/statutes/26/24/i.yaml with i.test.yaml. It should feed the 24.yaml credit for taxable years beginning after December 31, 2020 and before January 1, 2022.
  • Corpus citation: us/statute/26/24, subsection (i), 2021 edition (ARPA, Pub. L. 117-2, §9611).
  • Verbatim law: (i)(1)(B), (i)(3) and (i)(4)(A)-(D), as quoted under Summary.
  • Required outputs:
    • the applicable threshold amount ((i)(4)(B));
    • the applicable credit increase amount ((i)(4)(C)(ii));
    • the applicable phaseout threshold range ((i)(4)(C)(iii));
    • the (i)(4) reduction;
    • the credit after (i)(4) and (b);
    • the part allowed under subpart C under (i)(1)(B), and the remainder (the ODC) left in subpart A.
  • Companion tests from the 2021 Schedule 8812 and Line 5 Worksheet, not from PolicyEngine output. This is the same source as the PR's YAML. All three cases have one qualifying child under 6, and the first two add one other dependent:
    • HoH, MAGI $170,000. Worksheet line 11 is $1,600; Schedule 8812 line 5 is $2,000, line 12 is $2,500, 14a is $500 and 14b is $2,000.
    • HoH, MAGI $240,001. Line 11 is $2,050, line 12 is $450, 14a is $450 and 14b is $0.
    • Surviving spouse, MAGI $200,001, no other dependent. Worksheet line 6 is $2,500 and line 11 is $1,600; Schedule 8812 line 9 is $200,000, line 11 is $50 and line 12 and 14b are $1,950.
    • For h.test.yaml: a surviving-spouse case with a $200,000 threshold.
  • Why queued: the blocker, such as corpus scope, the signed release or approval for a billed run.

S2. No boundary or joint-return coverage of the new split

• Where:

  • T/refundable/refundable_ctc.yaml:216-327;
  • T/ctc.yaml:10-36;
  • policyengine_us/tests/policy/baseline/gov/states/ok/tax/income/credits/ok_federal_ctc.yaml:25.

• What the new cases cover.

  • F values. The cases put the first reduction F at $750 (single, $90,000) or at its $1,600 cap (HoH $170,000, HoH $240,001, surviving spouse $200,001).
  • No threshold case. None sits on an (i)(4) threshold. That is where F first becomes positive and where the old code first went wrong.
  • No joint return. Only ia_integration.yaml exercises the joint $150,000/$400,000 pair, with F equal to the $4,000 enhancement, P = 0 and no ODC. So neither phaseout is tested together with an ODC for a joint return.

• The Python grid doesn't fill the gap. It supplies ctc_arpa_phase_out and ctc_phase_out as inputs, so it never derives F or P from income at a threshold.

• Severity: should (missing boundary tests for changed logic).

• Fix: add household cases like these: 2021, TX, wages only, a child aged 5 and, where shown, a 10-year-old and a dependent aged 19. I computed the values from Schedule 8812 lines 5-14b and the Line 5 Worksheet. The last column shows what the old code would give.

  Case	F	P	ctc	refundable_ctc	non_refundable_ctc	Old code (refundable / nonrefundable)
  HoH $112,500, child 5 + dependent 19	0	0	4,100	3,600	500	same
  HoH $112,501, same family	50	0	4,050	3,550	500	3,600 / 450
  Single $75,001, same family	50	0	4,050	3,550	500	3,600 / 450
  Joint $150,001, children 5 and 10 + dependent 19	50	0	7,050	6,550	500	6,600 / 450
  Joint $400,001, same family	2,600	50	4,450	3,950	500	4,450 / 0
  Joint $480,001, same family	2,600	4,050	450	0	450	450 / 0

  • How the joint rows work. The enhancement is 3,600 + 3,000 − 4,000 = 2,600. That is below the 5% × 250,000 = 12,500 cap, so F is capped at 2,600 at both $400,001 and $480,001. P is 50 × 1 = 50 at $400,001 and 50 × 81 = 4,050 at $480,001.
  • The $112,500 row is the zero side of the boundary. The rows just above each threshold show the split moving by exactly $50.

S3. Iowa's 2021 add-back of non_refundable_ctc is now an obsolete workaround

• Where: policyengine_us/variables/gov/states/ia/tax/income/including_married_filing_separately/deductions/ia_fedtax_deduction.py:29-37. In fully refundable years it sets us_tax = ustax + non_refundable_ctc.
• History.
  • Why it was added. Commit 758955bfab (2023-05-05, "Fix ia_fedtax_deduction formula to work around 2021 US CTC flaw") added the add-back. At the time, the federal split gave the ia_integration.yaml family non_refundable_ctc −4,000 and refundable_ctc 12,000. Adding back the negative amount cancelled the error.
  • It is no longer needed. At the merge base the same fixture already shows 0 and 8,000, so the flaw it worked around is gone.
• What it does at this head. In 2021 non_refundable_ctc is now the ODC before the limit, min(ODC, line 12). income_tax_before_refundable_credits has already netted out the ODC actually applied. Iowa adds the full amount back:
  • When the ODC is fully used, Iowa's deduction ignores the ODC.
    • Below the ARPA thresholds this predates the PR: the old split also gave non_refundable_ctc = ODC.
    • Above them, the PR moves the overstatement out of the federal split and into this line.
  • When §26(a) limits the ODC, Iowa adds back more than was netted.
    • Trace. Take the PR's own Oklahoma household and put it in Iowa: 2021, single, $90,000, a child aged 5, a dependent aged 19, and income_tax_before_credits 300.
    • At this head. CI confirmed non_refundable_ctc 500 and income_tax_capped_non_refundable_credits 300 for that household. So income_tax_before_refundable_credits is 0 and us_tax = 0 + 500 = $500. That is more than the $300 of federal tax before any credit.
    • Before the PR, non_refundable_ctc was 0, so us_tax was $300.
    • The PR body is wrong here. It says the PR preserves Iowa's calculation, but this result changes.
• Iowa law.
  • The instructions. The 2021 IA 1040 instructions (file page 40) say "Report on line 31 all federal tax withheld, estimated tax payments, and federal taxes paid in 2021". The listed add-backs are Schedule 2 taxes. Nothing on pages 40-41 adds a nonrefundable credit back.
  • The formula's own premise. The formula approximates the payments as the liability: "With payments equal to the liability, the deduction is the federal income tax without the added-back taxes." The ODC reduces that liability, and the formula doesn't add it back in 2020 or 2022.
  • The comment. The new comment ("Retain the existing addback ... including any nonrefundable credit for other dependents in 2021") presents the workaround as an Iowa rule without a source.
• Severity: should. It's a wrong Iowa result in 2021, but in a state deduction the PR doesn't claim to change, and most of its reach predates the PR. That matches the Fix CalEITC child counts for ITIN holders #10322 S2 precedent.
• Fix: delete the fully refundable branch (lines 30-37), so that us_tax = ustax in every year.
  • Existing fixture. ia_integration.yaml doesn't move, because its non_refundable_ctc is 0.
  • New case. Add an Iowa 2021 case with an ODC (for example HoH, $170,000, a child aged 5 and a dependent aged 19). Assert non_refundable_ctc 500 and a head ia_fedtax_deduction equal to income_tax_before_refundable_credits less the Schedule 2 add-backs.
  • If you'd rather leave Iowa outputs alone in this PR, add back only the ODC actually applied (ctc_value − refundable_ctc, which is min(ODC, limiting tax)), reword the comment and file a follow-up issue.

Suggestions

G1. The Python grid test

• The cost is fine. policyengine_us/tests/test_ctc_2021_refundable_split.py ran in 1.42 s in CI (1.22 s setup).
• CLAUDE.md's "Test design and CI cost" rules still apply:
  • Duplicated formula. test_other_years_preserve_partial_refundability (lines 137-155) restates the legacy refundable_ctc formula and compares the code with itself. The existing refundable_ctc.yaml cases already cover those years.
  • YAML can do this. The 2021 partition checks could be vectorized YAML cases over tax units, with ctc_maximum, ctc_arpa_phase_out and ctc_phase_out as array inputs.
  • Hard-coded amounts. Line 60 writes 3_600/3_000 instead of reading amount.arpa.
  • No measurement note. The PR body doesn't give the measurement or the reason for using Python.
• Fix (optional): drop the other-years test or move its distinct cases to YAML. Otherwise, add a sentence to the PR body naming the failure only this grid detects, with the CI timing above.

G2. References and comment housekeeping

• V/phase_out/arpa/ctc_arpa_addition.py:11 still cites the current-year i1040s8.pdf#page=4, which no longer has the Line 5 Worksheet. The PR moved its two siblings (ctc_arpa_max_addition.py, ctc_arpa_phase_out_cap.py) to irs-prior/i1040s8--2021.pdf#page=4. Do the same here.
• T/refundable/refundable_ctc.yaml:210-211 has leftover scaffolding: "# Proposed additions to:" followed by the file's own path. Delete both lines.
• Lines 213-214 list the govinfo PDF twice, differing only in #page=. CLAUDE.md treats those as one source. Keep one href (#page=57) and say "PDF pages 57-58".

G3. The PR body is stale

• Draft language. It opens with "WIP: hub impact and parity are pending" although the PR is not a draft.
• Open items. Two items are still unchecked ("Finish validation" and "Hub").
• Stale test list. It lists 27 test files as "pending" or "interrupted". CI has since passed all of them (see CI status).
• Fix (optional): replace the validation section with the CI result, and remove the WIP line once S1 is settled.

Verified correct

• Hand-checks (8 cases, about 35 expectations). Unless noted, each is 2021, TX and wages only.
  1. T/refundable/refundable_ctc.yaml:216, HoH $170,000, a child aged 5 and a dependent aged 19.
     • F. Uncapped 50 × ceil(57,500 / 1,000) = 2,900. The caps are $1,600 (the enhancement) and $4,375 (5% × 87,500), so F = 1,600.
     • P = 0. The other expectations are ctc 2,500 (line 12), non_refundable_ctc 500 (14a), refundable_ctc 2,000 (14b) and ctc_value 2,500.
     • Old code: 2,500 refundable.
  2. Same file, line 252, HoH $240,001.
     • F = 1,600 (uncapped 6,400). P = 50 × ceil(40,001 / 1,000) = 2,050.
     • Line 12 = 2,500 − 2,050 = 450, so 14a is 450 and 14b is 0. ctc_value is 450.
  3. Same file, line 290, solar.
     • Tax before credits. Taxable income is 170,000 − 18,800 = 151,200. On the 2021 HoH schedule the tax is 1,420 + 4,800 + 7,073 + 15,564 = 28,857.
     • Solar credit. The potential is 26% × 150,000 = 39,000. The limit is 28,857 − 500 = 28,357, because Worksheet B is off in 2021 and the solar credit follows the ODC (Form 5695 Credit Limit Worksheet). The credit is 28,357.
     • income_tax = 28,857 − 500 − 28,357 − 2,000 = −2,000.
       • Recovery rebate. The 2021 HoH rebate is fully phased out by $120,000.
       • AMT. It doesn't apply: (170,000 − 73,600) × 26% = 25,064, which is less than 28,857.
     • Old code: −2,500, which is the PR body's $500 overstatement.
  4. T/ctc.yaml:10, surviving spouse, $200,001, a child aged 5.
     • F. The ARPA threshold is $150,000. F is uncapped at 50 × 51 = 2,550, capped at $2,500 and $1,600, so F = 1,600.
     • P. The threshold is $200,000, so P = 50. ctc and refundable_ctc are 1,950, and non_refundable_ctc is 0.
     • Old code: P = 0 and ctc 2,000.
  5. T/phase_out/arpa/ctc_arpa_phase_out_cap.yaml.
     • Joint with a $1,600 enhancement: min(12,500, 1,600) = 1,600.
     • Surviving spouse with a $5,000 enhancement: min(5% × (200,000 − 150,000), 5,000) = 2,500, with thresholds 150,000 and 200,000. This matches Line 5 Worksheet line 6.
  6. T/phase_out/ctc_phase_out_threshold.yaml.
     • 2017: [75,000, 110,000] (§24(b)(2)(B) and (A)).
     • 2018 and 2021: [200,000, 400,000], with ARPA thresholds [150,000, 150,000] in 2021.
     • 2026: [200,000, 400,000].
  7. ok_federal_ctc.yaml:25, single, $90,000, a child aged 5, a dependent aged 19, tax $300.
     • F = 50 × 15 = 750 (the caps are $1,600 and $6,250). P = 0.
     • ctc is 3,350, refundable_ctc 2,850 and non_refundable_ctc 500. Capped nonrefundable credits are min(300, 500) = 300.
     • ctc_value = min(3,350, 300 + 2,850) = 3,150. ok_federal_ctc = 300 applied + 2,850 = 3,150 (Credit Limit Worksheet A, binding branch).
  8. ia_integration.yaml (comments only), joint, $252,020, four children aged 11, 11, 11 and 16.
     • C = 12,000, the base is 8,000 and the enhancement 4,000.
     • F. Uncapped 50 × 103 = 5,150, capped at 4,000. P = 0.
     • ctc and refundable_ctc are 8,000, and non_refundable_ctc is 0. That is unchanged.
• The Python grid. Its 2021 assertions are lines 12, 14a and 14b, plus refundable + non_refundable == ctc. All 7 tests pass in CI.
• Callers at the head:
  • Solar limit. residential_clean_energy_credit_credit_limit now counts only the ODC as preceding in 2021.
  • Oklahoma. ok_federal_ctc applies the capped amount.
  • Missouri. mo_federal_income_tax_deduction adds back refundable_ctc (line 28) to income_tax. With the ODC now on line 19, the base nets it, as the MO-1040 line 9 worksheet requires.
  • Alabama. The 2021 Part II of al_federal_income_tax_deduction adds back ctc − refundable_ctc and subtracts the recomputed 2020 amount. In the non-binding cases the result is unchanged. In the binding cases I traced, the line-6 zero floor absorbs the difference.
  • New York. The ny_ctc_pre_2024 branch pins every CTC parameter to 2017, so neither change reaches it.
  • Puerto Rico and eligibility. pr_refundable_ctc and eligible_for_refundable_credits read the corrected amounts.
  • Iowa is S3.
  • Threshold readers. The other readers of ctc_phase_out_threshold are contrib reforms, taxsim.py and pr_ctc_phase_out, which follow the corrected surviving-spouse value.
• References.
  • The govinfo PDF. File pages 57-58 hold §24(h)-(i); page 57 prints "Page 80". File page 81 holds §25D(c).
  • The forms. f1040s8--2021.pdf#page=1 shows line 9, and i1040s8--2021.pdf#page=4 shows the Line 5 Worksheet.
  • Format. The multi-page govinfo href in V/refundable/refundable_ctc.py carries a "PDF pages 57-58" comment, as CLAUDE.md asks. The new threshold reference title ("2021 Schedule 8812, line 9") has no page text.
• Changelog. changelog.d/ctc-2021-refundable-split.fixed.md is a top-level fragment, and fixed is the right type.
• Out of scope (pre-existing, not findings):
  • Part I-C. The model gives every 2021 filer the fully refundable path. It doesn't model filers without a US principal place of abode (Part I-C), which Remove calculated variables from input/income.py #1829 tracks on the rulespec side.
  • Advance payments. It doesn't model the §24(j)/§7527A advance payments. On the form, lines 14f-14i take advances from the refundable part first, but PolicyEngine models the annual entitlement, so the 14a/14b split is the right target.
• Batch lessons:
  • Source-conflict ratchet (Tie source-conflict markers to the register shared with Axiom #10287). The scanned roots are variables, parameters and reforms.
    • The check. I ran main's CONFLICT_COMMENT_RE over the comments of the six scanned files the PR touches (threshold.yaml, non_refundable_ctc.py, ctc_arpa_max_addition.py, ctc_arpa_phase_out_cap.py, refundable_ctc.py and ia_fedtax_deduction.py) at the merge base, the head and main. Each has 0 matches.
    • The removed Iowa comment ("flaw in CTC refundable/nonrefundable split") doesn't match.
    • The baseline list. None of these files is in KNOWN_UNMARKED, and the PR adds no marker.
    • Result. The guard passes on the merged tree. The test YAML and the new Python file are outside the scan.
  • NH Add New Hampshire state supplementation for 2026 #10156 (state-gated keys): not applicable. Both thresholds are filing-status breakdowns with all five keys, including SURVIVING_SPOUSE.
  • Integrate modeled state unemployment insurance into income flows #9558, Document disjoint health premium inputs and correct consumer treatment #10046 and OK Add Oklahoma state supplementation for 2026 #10165: not applicable. No income-source list or wage definition changes.
  • Partner tests: untouched. The Household API Partners check passed.

Axiom line assessment

Invalid: axiom: pending: hub files parity (S1).

• Recommended form: axiom: us:statutes/26/24/h#ctc_phase_out_threshold_under_subsection_h encoded-correct; TheAxiomFoundation/rulespec-us#<n> queued. The queued issue is a new open, pe-parity-labelled, dispatch-ready issue for §24(i)(1)(B), (i)(3) and (i)(4). The required contents are listed under S1. Alternatively, cite an encoder PR as encoded.
• Not n/a. These are statutory provisions, and the PR changes 2021 refunds and 2018+ surviving-spouse credits.
• Not encoded-correct for §24(i). rulespec-us main (5796861e58) has no subsection (i) module. Remove calculated variables from input/income.py #1829 covers only (i)(1)(A).
• encoded-correct is available for §24(h)(3), provided filing_status == 1 means a joint return only. Its tests lack a surviving-spouse case, so add one in the new issue.

CI status

27 pass, 3 skipping, 0 pending, 0 failed. I re-checked gh pr checks 10306 at the end of this review, and the head is still cffa308032.

• The green run. Run 39515880371 (pull_request, created 2026-10-20 23:34 UTC) tested merge commit afedcfae. That merges the head into 07eba1aeea, which is main as of 2026-10-20 19:22 UTC. That is before Tie source-conflict markers to the register shared with Axiom #10287 (22:02 UTC) and Count the FL and SD LIHEAP comments in the source-conflict ratchet baseline #10334, so CI's tree did not contain the ratchet. I checked it statically instead (see Verified correct).
• Where the changed tests ran.
  • Quick Feedback selected and passed: ctc.yaml (2), ctc_arpa_phase_out_cap.yaml (3), ctc_phase_out_threshold.yaml (6), refundable_ctc.yaml (18), ia_integration.yaml (9) and ok_federal_ctc.yaml (6). It also passed the new Python file (7 tests in 1.42 s).
  • Rest (Python, shard-1 + variables) passed the Python file again.
  • The full baseline, contrib, microsimulation and Household API Partners suites passed. That includes every file the PR body lists as unfinished.
• The skipped checks are the umbrella jobs of run 39489870440 (2026-10-20 20:11 UTC), which the green run superseded.
• Locally: I ran no tests or model probes. CI is green at the current head. Every expectation and the S3 trace were checked by hand from the code, the parameters and CI-verified values.

Branch status

• Position: 4 commits ahead of main (328d1721e2) and 26 behind. The merge base is f187dae556.
• Merge: clean.
  • git merge-tree --write-tree PolicyEngine/main cffa308032 exits 0 (tree 4df778bf2e).
  • Since the merge base, main has touched Iowa itemized deductions, Oklahoma LIHEAP, SSP and additions, and some itemized-deduction tests. It has touched no file in this PR and nothing under the federal CTC.
• CI base (note only). CI's base is 20 commits behind main. Merging main and re-running would refresh CI on a tree that contains the ratchet guard.
• Siblings (note only). No other open PR edits refundable_ctc.py, the CTC phaseout parameters, non_refundable_ctc.py, ia_fedtax_deduction.py or ok_federal_ctc.py.

0 critical, 3 should and 3 suggestions are open at this head.

Verdict: REQUEST_CHANGES

[DTrim99]

The 2021 split now matches s.24(i)(1)(B) and (i)(4) and Schedule 8812 lines 12/14a/14b, the surviving-spouse $200k threshold matches s.24(h)(3), and other years are unchanged. Three should-address items: (1) the 'pending: hub files parity' axiom line is invalid, so please file a queued pe-parity issue for 24(i)(1)(B)/(i)(3)/(i)(4); (2) no new case sits on an ARPA threshold or covers both phase-outs plus the ODC (the report gives the values); (3) Iowa's 2021 non_refundable_ctc add-back is an obsolete workaround that now over-adds ($500 vs $300). Three suggestions. Full report in the comment above.

[MaxGhenis]

Evidence from Tax-Calculator's PE-TAXSIM validation (PSLmodels/Tax-Calculator taxcalc/validation/pe_taxsim, 10,000 synthetic units per sample), run at this PR's head cffa308 through policyengine-taxsim 3.2.2 and compared with Tax-Calculator 6.8.5 (|tc − pe| > $1):

2021 sample	actc differences, PE-US 2.58.0	at cffa308
a21 (wages, ages)	1,489	8
b21 (+ other income)	1,009	4
c21 (+ deductions, child care)	1,082	6

fiitax and v22 agree before and after; only the refundable/nonrefundable split moves, by min(ODC, ARPA reduction) per unit where no §24(b) reduction applies. 2022–2025 samples are unchanged.

Of the 18 remaining units, 11 are new (PE-US 2.58.0 and Tax-Calculator agreed on a split that is wrong under the form) and 7 differed before. All 18 are heads of household with AGI $328,000–$406,099, where the §24(b) reduction exceeds the $2,000-per-child base. Tax-Calculator phases out its base CTC plus ODC before the ARPA increase (ctc_new), so the reduction there falls on the ODC first. This PR follows Schedule 8812 line 14a. Example, a21 unit 3349 (HoH, AGI $336,000, children 1, 2 and 4, one dependent aged 22): Line 5 Worksheet $10,800 − $4,375 = $6,425; line 8 $6,925; line 11 $6,800; line 12 $125; line 14a min($500, $125) = $125; line 14b $0. This PR: refundable $0. Tax-Calculator: $125.

The TAXSIM-35 test binary (cd2026081819) reports actc equal to v22 for all 18 units, so this PR moves the emulator's 2021 actc away from TAXSIM-35 and onto the form.

[MaxGhenis]

Picking this up for the US hub. I'll post a hand-off comment here when it's ready.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review

Head: 0a9022de26071d3974587991d53555418a650fda (fast-forward from cffa308032).

What changed

• Merge main (c5886d5946). One conflict, in refundable_ctc.yaml; I kept both sides' cases.
• Should 2, boundary and joint cases (8e8edf6e60): six 2021 TX wage-only households in refundable_ctc.yaml:364-597. Each has a child aged 5 (plus a child aged 10 on joint returns) and a dependent aged 19 who is not a student, with explicit head, spouse and dependent roles. Without those roles the single-parent units did not count the 19-year-old.
• Should 3, Iowa (8a75fd7070): ia_fedtax_deduction.py:31 now sets us_tax = income_tax_before_refundable_credits in every year. The 2021 non_refundable_ctc add-back is gone, and the comment says the 2021 ODC reduces the liability paid. New case at ia_fedtax_deduction.yaml:127. The changelog fragment gains an Iowa line.
• Suggestions:
  • G2 is taken (ctc_arpa_addition.py:11 and the YAML scaffolding).
  • G1 is taken in part (0a9022de26): the grid reads amount.arpa.
  • G3 is in the PR body.

Hand computations (2021 Schedule 8812 lines 5-14b, Line 5 Worksheet; 26 U.S.C. 24(i)(4), (b), (h))

Child amounts are $3,600 (under 6) and $3,000 (6-17), against a $2,000 base, with a $500 ODC. F is the first reduction: $50 per $1,000 or fraction above the (i)(4)(B) threshold, capped at the lesser of the enhancement and 5% of (the (h)(3) threshold − the (i)(4) threshold). P is the ordinary phaseout above $200,000 / $400,000.

Case	F	P	ctc (line 12)	refundable_ctc (14b)	non_refundable_ctc (14a)
HoH $112,500	0	0	3,600 + 500 = 4,100	3,600	500
HoH $112,501	50 × ceil(1/1,000) = 50 (caps 1,600; 4,375)	0	4,050	3,550	500
Single $75,001 (status supplied)	50 (caps 1,600; 6,250)	0	4,050	3,550	500
Joint $150,001	50 (caps 2,600 = 6,600 − 4,000; 12,500)	0	7,100 − 50 = 7,050	6,600 − 50 = 6,550	500
Joint $400,001	min(50 × 251 = 12,550; 2,600; 12,500) = 2,600	50 × ceil(1/1,000) = 50	7,100 − 2,650 = 4,450	6,600 − 2,650 = 3,950	500
Joint $480,001	2,600	50 × ceil(80,001/1,000) = 4,050	7,100 − 6,650 = 450	max(0, 6,600 − 6,650) = 0	min(500, 450) = 450

These match the review's table.

Iowa (2021 HoH, wages $170,000, a child aged 5 and a dependent aged 19):

• Federal taxable income: 170,000 − 18,800 = 151,200.
• Tax: 1,420 + 4,800 + 7,073 + 24% × (151,200 − 86,350) = 28,857.
• The $500 ODC brings income_tax_before_refundable_credits to 28,357. There are no Schedule 2 add-backs, and the head's prorate fraction is 1, so ia_fedtax_deduction = [28,357, 0, 0]. The old add-back gave 28,857.

Tests (local, one file per process)

• YAML. refundable_ctc.yaml 26, ctc.yaml 2, ctc_arpa_phase_out_cap.yaml 3, ctc_phase_out_threshold.yaml 6, ia_fedtax_deduction.yaml 5, ia_integration.yaml 9 (unchanged expectations), ok_federal_ctc.yaml 6. All pass: 57.
• Python. test_ctc_2021_refundable_split.py 7, test_source_conflicts.py 103. Both pass.
• Mutants, each in its own process, restored after:
  • Restoring the old 2021 split (refundable = child amounts − P only) fails 8 of 26 in refundable_ctc.yaml: five of the six new cases (the $112,500 case is the intended no-change side) plus the PR's three.
  • Restoring Iowa's add-back fails the new Iowa case (28,857 vs 28,357; 1 failed of 5).

CI

Pull request run 41730727280 at 0a9022de26: completed, success, all 36 jobs passed. The tree merges current main, including the #10287 ratchet.

Judgment calls

• No source was reachable from this VM (the proxy refused irs.gov, govinfo and revenue.iowa.gov). The statute and form steps above follow the PR's references and the review's quotations, unverified by me. Iowa: I relied on the code's own premise (payments equal the liability) and the review's reading of the 2021 IA 1040 instructions (federal tax paid, Schedule 2 add-backs only, no nonrefundable credit added back). I could not read those instructions.
• Iowa: removed, not narrowed. The review's alternative, adding back only the ODC actually applied, would keep a workaround with no Iowa source.
• G1, the other-years grid test, is kept. It checks that the 2021 branch doesn't leak into 2018, 2020, 2022, 2025 and 2026 across a grid. That's coverage the year-specific YAML doesn't give, and the file runs in about 2 s. Only its hard-coded amounts changed.

Reply to DTrim99

1. S1 (axiom line). Left to the hub, as instructed. The body's axiom: line is unchanged.
2. S2 (boundary and joint coverage). Taken. Six cases at refundable_ctc.yaml:364 (HoH $112,500), :397 (HoH $112,501), :431 (single $75,001), :465 (joint $150,001), :510 (joint $400,001) and :555 (joint $480,001). Each asserts AGI, ctc_arpa_phase_out, ctc_phase_out, ctc, refundable_ctc and non_refundable_ctc, with the values in your table. The old split fails five of them.
3. S3 (Iowa). Taken: the fully refundable branch is deleted (ia_fedtax_deduction.py:31), so us_tax is income_tax_before_refundable_credits in every year. ia_integration.yaml doesn't move. The new case (ia_fedtax_deduction.yaml:127, HoH $170,000 with a child and a dependent aged 19) asserts non_refundable_ctc 500, federal tax 28,357 and a head deduction of 28,357. The fragment and the body no longer say Iowa's calculation is preserved.
4. G1 (Python grid). Taken in part: line 60 now reads amount.arpa (test_ctc_2021_refundable_split.py:60). I kept test_other_years_preserve_partial_refundability. It's the only check that the 2021 branch leaves five other years unchanged on a grid. It restates the legacy formula, but cheaply (about 2 s locally). The body now names what the grid detects.
5. G2 (references and housekeeping). Taken:
   • ctc_arpa_addition.py:11 cites irs-prior/i1040s8--2021.pdf#page=4;
   • the "Proposed additions to:" scaffolding is gone;
   • the govinfo PDF is cited once, with a "PDF pages 57-58" note (refundable_ctc.yaml:210-212).
6. G3 (PR body). Taken: the WIP line is gone, the stale validation list is replaced with the test and CI results, and the finished plan items are ticked (the hub's impact and parity items stay open).

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

This review is finished. The full review is in #10306 (comment).

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: APPROVE
0a9022d
METHODOLOGY_CHOICE: yes

Independent review, round 2. The head is still 0a9022de26, the hand-off's head, so nothing changed after it.

Methodology choice: Iowa. The PR removes the 2021 add-back rather than adding back only the ODC actually applied. The two choices give different Iowa deductions. Removal deducts the federal tax after nonrefundable credits, which is what the formula already does in every other year. I couldn't read the Iowa sources to confirm it.

The surviving-spouse threshold isn't a methodology choice. §24(h)(3) as quoted in the first review, and Schedule 8812 line 9, both give the $200,000 threshold.

Findings

P2-1. A stale Iowa test comment says the removed add-back still exists

• File: policyengine_us/tests/policy/baseline/gov/states/ia/tax/income/ia_integration.yaml:56
• Failure scenario: the comment reads # Iowa retains its addback of nonrefundable CTC in fully refundable years. Commit 8a75fd7070 deleted that add-back. ia_fedtax_deduction.py:31 now sets us_tax = income_tax_before_refundable_credits in every year. A reader of the fixture would take the reverse of the code's behavior. The hand-off says the fragment and the body no longer claim Iowa is preserved, but this comment still does.
• Command:

  grep -n "retains its addback" -r policyengine_us/tests/policy/baseline/gov/states/ia
  

  Output:

  ia_integration.yaml:56:    # Iowa retains its addback of nonrefundable CTC in fully refundable years.
  

• Replacement:

      # Iowa deducts income_tax_before_refundable_credits with no CTC addback; non_refundable_ctc is 0 here.
  

P2-2. The body has no impact: line, and the diff moves baseline outputs

The body has an axiom: line but no impact: line. The plan item "Hub: run the impact analysis" is still open. The diff moves outputs at baseline. Probe runs at origin/main (a61a874778) and at this head, using the same script, show four effects:

Household	main	Head
2021 HoH, $112,501, child aged 5 and a dependent aged 19: refundable / nonrefundable CTC	3,600 / 450	3,550 / 500
Same household: income_tax_before_refundable_credits	14,607.24	14,557.24
2021 joint, $400,001, children aged 5 and 10 and a dependent aged 19: refundable / nonrefundable CTC	4,450 / 0	3,950 / 500
2021 surviving spouse, $250,000, child aged 5: income_tax	40,018	42,018
2021 HoH, $170,000, Iowa: ia_fedtax_deduction (head)	28,857	28,357

• The 2021 split. Total federal tax moves only where the ODC or later nonrefundable credits meet the §26(a) limit. The solar case is one example.
• The surviving-spouse threshold. It applies every year from 2018, including current years. It cuts the CTC for surviving spouses with qualifying children and MAGI above $200,000.
• Iowa 2021. The deduction falls by non_refundable_ctc for every Iowa tax unit with an ODC, so Iowa tax rises.

The population check therefore needs:

• 2021 federal income_tax and refundable_ctc;
• the CTC for surviving-spouse units in 2018 through 2026;
• 2021 Iowa income tax.

The impact run is the hub's job. I didn't run it.

P3-1 (follow-up, the same on main). The 2021 branch skips the §911 bar

refundable_ctc.py:34-43 returns inside if ctc.refundable.fully_refundable before it reads refundable_ctc_barred_by_section_911_exclusion. So a 2021 filer who elects the §911 exclusion still gets a fully refundable child amount.

Witness, on main at a61a874778. The inputs are ctc 3,600, ctc_refundable_maximum 3,600, both reductions 0, and refundable_ctc_barred_by_section_911_exclusion: true:

refundable_ctc@2021: [3600.] differs from 0.0

This PR didn't introduce it. Whether §24(d)(3) or Schedule 8812 Part I-C bars these filers in 2021 needs the 2021 text, which I couldn't reach from this VM. So it's a follow-up candidate, not a finding against this PR. The first review's "Part I-C" note covers the same gap.

Checked and sound

• Should 2: the six boundary and joint cases (refundable_ctc.yaml:364, 397, 431, 465, 510, 555).

  • By hand first. I recomputed each case before reading the hand-off table:

    • F = min(50 × ceil(excess over the §24(i)(4)(B) threshold / 1,000), the enhancement, 5% × (the §24(h)(3) threshold − the (i)(4) threshold)).
    • P = 50 × ceil(excess over $200,000 or $400,000 / 1,000).
    • Line 12 = C − F + O − P, 14b = max(C − F − P, 0), and 14a = line 12 − 14b.

    Case	F	P	ctc	refundable_ctc	non_refundable_ctc
    HoH $112,500	0	0	4,100	3,600	500
    HoH $112,501	50 (caps 1,600 and 4,375)	0	4,050	3,550	500
    Single $75,001	50 (caps 1,600 and 6,250)	0	4,050	3,550	500
    Joint $150,001	50 (caps 2,600 and 12,500)	0	7,050	6,550	500
    Joint $400,001	min(12,550, 2,600, 12,500) = 2,600	50	4,450	3,950	500
    Joint $480,001	2,600	50 × 81 = 4,050	450	0	450

    All match the YAML.

  • Roles. A per-person probe at the head, for HoH $112,500 and joint $400,001, shows:

    • The dependent aged 19: ctc_qualifying_child False, ctc_child_individual_maximum_arpa 0, ctc_adult_individual_maximum 500. So they count for the $500 ODC only.
    • The children aged 5 and 10: 3,600 and 3,000.

  • Strength. Each case asserts F, P and all three CTC values, so none asserts a weaker claim. The +$1 cases sit on the boundary from above, and the $112,500 case is the zero side.

• Should 3: Iowa (ia_fedtax_deduction.py:31; case at ia_fedtax_deduction.yaml:127).

  • The parameters. I printed the 2021 HoH brackets from the repo's parameters: 14,200 / 54,200 / 86,350 / 164,900 at 10 / 12 / 22 / 24%. The standard deduction is 18,800.
  • Federal tax. (170,000 − 18,800) = 151,200 of taxable income. The tax is 1,420 + 4,800 + 7,073 + 0.24 × 64,850 (15,564) = 28,857.
  • After the ODC: 28,857 − 500 = 28,357. The head's deduction is [28,357, 0, 0].
  • Reach. The deleted branch only differed when fully_refundable is true. That is 2021 at baseline. The two reforms that set it in later years (afa_other_dependent_credit, wyden_smith/ctc_expansion) replace refundable_ctc with their own. Iowa's §26(a)-limited case now follows automatically from income_tax_before_refundable_credits.

• Suggestions.

  • G1: test_ctc_2021_refundable_split.py:60 reads p.amount.arpa.calc(...). The other-years test is kept, with the reason in the body. It runs in 1.37 s locally (1.32 s of setup).
  • G2:
    • ctc_arpa_addition.py:11 cites irs-prior/i1040s8--2021.pdf#page=4;
    • the "Proposed additions to:" scaffolding is gone;
    • refundable_ctc.yaml:210-213 cites the govinfo PDF once, with "PDF pages 57-58".
  • G3: the body has no WIP line, and every plan item is ticked except the hub's.

• The merge of main (c5886d5946, parents cffa308032 and 6145ad4513).

  • Cases. I loaded refundable_ctc.yaml at both parents and at the merge with PyYAML. All 18 branch-side cases and all 17 main-side cases appear unchanged in the merge's 20 (the two sides share 15).
  • Other files. git diff <main parent> <merge> carries exactly the changed lines of git diff <merge base> cffa308032, so the merge adds nothing else.

• The code against the rule. In 2021, refundable_ctc = min(max(0, C − F − P), ctc). Since ctc ≥ that amount whenever C equals the child amounts, this is line 14b, and non_refundable_ctc = ctc − R is min(O, line 12) = 14a. Other years keep the legacy branch, and the grid's five other-year tests pass.

• Defects reproduced on main. These are the "Reproduce on main" rows in P2-2: the old split refunds up to F of the ODC, and the surviving-spouse threshold is 400,000.

• Tests at the head, one file per process. All pass:

  File	Result
  refundable_ctc.yaml	26
  ctc.yaml	2
  ctc_arpa_phase_out_cap.yaml	3
  ctc_phase_out_threshold.yaml	6
  ia_fedtax_deduction.yaml	5
  ia_integration.yaml	9
  ok_federal_ctc.yaml	6
  test_ctc_2021_refundable_split.py	7
  test_source_conflicts.py	103

• Mutants. Each ran in its own process, and the tree was restored after each (git status clean, HEAD 0a9022de26):

  Mutant	Result	Killed by
  M1 (hand-off): 2021 reduction = ctc_phase_out only	8 of 26 fail in refundable_ctc.yaml; 2 of 7 in the grid	The PR's three cases (refundable 2,500 vs 2,000; 450 vs 0; solar limit 28,857 vs 28,357) and five of the six new cases. All but $112,500, which is the intended no-change side.
  M2 (hand-off): restore Iowa's add-back	1 of 5 fails	ia_fedtax_deduction.yaml:127 (28,857 vs 28,357). ia_integration.yaml still passes 9.
  M3 (mine): surviving-spouse threshold back to 400,000	5 fail	ctc.yaml (1), ctc_phase_out_threshold.yaml (the 2018, 2021 and 2026 cases) and the surviving-spouse case in ctc_arpa_phase_out_cap.yaml
  M4 (mine): use ctc_arpa_uncapped_phase_out in the split	3 of 26 fail	HoH $170,000 (700 vs 2,000), solar (27,057 vs 28,357) and joint $400,001 (0 vs 3,950)
  M5 (mine): ceil → floor in ctc_arpa_uncapped_phase_out	3 of 26 fail	Exactly the three new +$1 cases (HoH $112,501, single $75,001, joint $150,001). No earlier case caught it.

• Readers of refundable_ctc, non_refundable_ctc, ctc and ctc_phase_out_threshold:

  • Federal readers:
    • ctc_value;
    • residential_clean_energy_credit_credit_limit (tested);
    • eligible_for_refundable_credits;
    • the refundable.yaml and non_refundable.yaml lists.
  • States:
    • ok_federal_ctc (tested; unchanged at 2,500 for HoH $170,000);
    • al_federal_income_tax_deduction (12,557.24 and 77,172.32, the same on main and the head);
    • mo_federal_income_tax_deduction (0 on both for these incomes);
    • md_ctc, which reads federal ctc only within its low-AGI eligibility;
    • ny_ctc_pre_2024, which pins 2017 parameters.
  • Territories: pr_refundable_ctc and pr_ctc_phase_out.
  • Reforms:
    • pa_ctc_match, crfb/non_refundable_ss_credit, edaa_end_child_poverty_act, fisc_act, american_worker_rebate_act and ctc_minimum_refundable_amount;
    • taxsim and ctc_per_child_phase_out. Both redefine the ARPA reduction, which the new branch now subtracts consistently.

  I found no reader that this PR makes reachable and worse.

• Docs.

  • non_refundable_ctc documentation. It says "before the section 26(a) tax-liability limit" and "line 14a … before the limit on line 14d". Both match ctc − refundable_ctc in every year.
  • Reference format. The refundable_ctc.py multi-page href has its "PDF pages 57-58" comment, as CLAUDE.md asks. The new threshold reference has no page text in its title.
  • The changelog is one top-level fixed fragment.

• The body's claims checked:

  • the T, R and N formulas;
  • "five of the six new boundary cases" (M1);
  • the Iowa mutant (M2);
  • the test counts;
  • the run time ("about 2 s"; 1.37 s here);
  • the $500 solar overstatement (M1 shows 28,857 vs 28,357).

• CI. Pull request run 41730727280 at 0a9022de26 completed with success. All 36 jobs passed: baseline, contrib, Python, microsimulation, Household API Partners, PR validation and Quick Feedback. None was skipped or cancelled. The run's base is 6145ad4513; main is now 5 commits ahead, and none of those commits touches these files.

• Scope.

  • Nothing under .github/ or tests/policy/baseline/partners/.
  • Open-PR searches for refundable_ctc, non_refundable_ctc, ia_fedtax_deduction, threshold.yaml, ok_federal_ctc and ctc_arpa returned Reduce Oregon's federal tax subtraction by the OR-17 worksheet's refundable credits only #10093, Ignore a dependent's separation when inferring the tax unit spouse (26 U.S.C. 7703) #9932, Apply New Jersey, North Dakota, Ohio, Oklahoma and Pennsylvania rules for a filer who can be claimed as a dependent #10043, Model the claim of right repayment deduction and credit (26 U.S.C. 1341) and North Carolina's rules that depend on it #10038, Add Arkansas's claim of right deduction (Ark. Code § 26-51-457) #10203 and Add Indiana's, Massachusetts's and West Virginia's claim of right repayment deductions #10196. I listed their files, and none touches a file this PR changes. Apply New Jersey, North Dakota, Ohio, Oklahoma and Pennsylvania rules for a filer who can be claimed as a dependent #10043 edits nj_ctc and an Oklahoma EITC file only.

What I couldn't check

• Primary sources. govinfo.gov, uscode.house.gov, irs.gov, revenue.iowa.gov and legis.iowa.gov all refused connections from this VM (proxy 403 and DNS denial). The statute and Schedule 8812 steps above rest on the first review's quotations of 26 U.S.C. 24(b), (h) and (i) and of the 2021 Schedule 8812 and Line 5 Worksheet, not on my own reading. I couldn't read the Iowa Code section or the 2021 IA 1040 instructions. The new Iowa comment ("the 2021 credit for other dependents … reduces the liability paid") rests only on the formula's existing payments-equal-liability premise.
• The 2021 AMT. I confirmed the HoH exemption ($73,600) from the repo's parameters. I didn't confirm the 26% rate, which only shows that AMT doesn't bind in the Iowa and solar cases. The model's run agrees.
• Parity. I didn't check whether 2021 child tax credit: refundable child portion after both phaseouts, nonrefundable other-dependent credit (26 USC 24(i)(1)(B), (i)(3), (i)(4)) (PE-US #10306) TheAxiomFoundation/rulespec-us#1856 exists, is open and is labelled pe-parity. That repository is outside this session's access.
• Population impact. I ran no microsimulation (P2-2).

---

Generated by Claude Code