Fix the 2020 unemployment compensation exclusion: cutoff test, per-spouse cap, dependents

[MaxGhenis]

Fixes three gaps in the 2020 federal unemployment compensation exclusion (IRC 85(c), added by ARPA § 9042).

The law

26 U.S.C. 85(c) (checked 2026-10-12):

(1) In general. In the case of any taxable year beginning in 2020, if the adjusted gross income of the taxpayer for such taxable year is less than $150,000, the gross income of such taxpayer shall not include so much of the unemployment compensation received by such taxpayer (or, in the case of a joint return, received by each spouse) as does not exceed $10,200.
(2) Application. For purposes of paragraph (1), the adjusted gross income of the taxpayer shall be determined— (A) after application of sections 86, 135, 137, 219, 221, 222, and 469, and (B) without regard to this section.

Section 85(a) is what puts unemployment compensation in gross income, so "without regard to this section" leaves all of it out. The IRS implements this in the Unemployment Compensation Exclusion Worksheet in the 2020 Instructions for Form 1040 (rev. 4-2021), file page 88:

• Lines 1–4 add total income without unemployment compensation (line 2: "Don't include any amount of unemployment compensation from Schedule 1, line 7").
• Line 5 subtracts the adjustments, and line 6 is "your modified adjusted gross income".
• Line 7: "$150,000 or more? Yes. Stop."
• Lines 8 and 9 take the compensation paid to you and, on a joint return, to your spouse, each up to $10,200.

Page 87 has the joint example: "$20,000 … and your spouse was paid $5,000 … report $15,200 on line 8 as a negative amount".

What was wrong

1. The compensation was subtracted twice in the $150,000 test. taxable_uc_agi already sums the gross income sources other than taxable_unemployment_compensation, less above-the-line deductions, so it is AGI without the compensation (worksheet line 6). tax_unit_taxable_unemployment_compensation then compared taxable_uc_agi - tax_unit_unemployment_compensation with the cutoff. A single filer in Montana, age 66, with $150,000 of taxable pension income and $8,000 of compensation, got taxable_unemployment_compensation 0. The law gives 8,000, because line 6 is 150,000, which is not under $150,000. The two boundary tests added in Fix 2020 unemployment exclusion boundary at $150,000 AGI #8900 encoded the same misreading: their comments say "AGI over UC = 155,000 - 5,000 = 150,000".
2. The $10,200 was per tax unit, not per spouse. The model excluded min(tax unit compensation, 10_200). In the IRS joint example it gave taxable [11_840, 2_960] (14,800 in all); the law gives [9_800, 0].
3. Dependents' compensation used the filers' exclusion. tax_unit_unemployment_compensation adds every member, dependents included. Take a 2020 joint couple whose spouse received $8,000 and whose 20-year-old student dependent received $4,000. The model gave taxable [0, 1_200, 600], and irs_gross_income then dropped the dependent's 600. The law gives the couple 0 taxable: the spouse's 8,000 is under 10,200, and the dependent's 4,000 is on the dependent's own return.

Changes

• taxable_unemployment_compensation (Person) is now each person's own compensation less their own exclusion. That exclusion is min(max(0, compensation), 10_200) when the AGI of the return the compensation is on is under the cutoff, and nothing otherwise.
  • Head and spouse: taxable_uc_agi of the tax unit, compared directly with the cutoff, with no second subtraction.
  • Dependent: the new dependent_taxable_uc_agi, their own § 85(c)(2) AGI, against the single cutoff, or the separate one if they live with their spouse. It reuses dependent_income_and_deductions from dependent_taxable_ss_magi.py, less the compensation, plus dependent_taxable_social_security.
  • Taxable compensation is never negative. irs_gross_income counts positive amounts only, so a net repayment never reduced gross income. Before this PR a person's share was 0 when the tax-unit total was not positive. A negative taxable amount would let California, Pennsylvania and Massachusetts subtract it. tax_exempt_unemployment_compensation (received less taxable) carries the repayment as a negative amount, as before this PR, which Wisconsin's homestead income reads.
• tax_unit_taxable_unemployment_compensation is the sum over members who are not dependents (sum_among_non_dependents), as irs_gross_income counts them.
• taxable_uc_agi: the label, documentation and references now say what it is (worksheet line 6). The formula is unchanged.
• The exemption amount and cutoff parameters get descriptions, labels and references (IRC 85(c), worksheet lines).
• tax_unit_unemployment_compensation is unchanged and no longer feeds the exclusion.
• dependent_taxable_ss_magi.py: estate_income joins BUSINESS_LOSS_SOURCES, so a dependent's own estate or trust loss (Schedule E Part III, Schedule 1 line 5) reduces the modified AGI of the dependent's own return. limited_business_loss already deducts these losses for the head and spouse, and the list's own comment says it holds the sources loss_ald deducts. The shared helper feeds dependent_taxable_uc_agi, dependent_taxable_ss_magi (and so dependent_taxable_social_security) and dependent_traditional_ira_deduction, so all three now net the loss. Tests: dependent_taxable_uc_agi.yaml Case 3 and a new dependent_taxable_ss_magi.yaml case.
• The same helper now applies IRC 461(l) on the dependent's own return, as limited_business_loss does for the filers. Business losses (self-employment, farm, rental, partnership/S corporation, estate) are allowed up to business income plus gov.irs.ald.loss.max for a single (or, living with a spouse, separate) filer. That threshold is infinite for 2018–2020 (CARES Act § 2304), so the 2020 test keeps the whole loss. Without the limit, a $700,000 estate business loss would have wiped out a 2025 dependent's taxable Social Security. Tests: the 2025 above-limit case in dependent_taxable_ss_magi.yaml (modified AGI 88,000, taxable benefits 17,000), and dependent_taxable_uc_agi.yaml Case 4 (2020, suspended).

Direction of allocation. Before, the person-level amount was the tax-unit amount times the person's share of the unit's compensation. Now persons are primary and the tax unit sums them. No dataset input sets tax_unit_taxable_unemployment_compensation: the hub's screen sample of the default dataset, which keeps its inputs, has 18 tax-unit columns and none is this one, and no reform overrides it. One test sets it as an input (in_unemployment_compensation_deduction.yaml Case 1), and only Indiana's deduction reads it there, which is unchanged. The only remaining readers of the tax-unit variable are Indiana's deduction and taxsim_v11.

Effect outside 2020

The exclusion amount is 0 outside 2020, so the person-level taxable_unemployment_compensation is each person's own compensation in every other year, before and after this PR. The tax-unit total now leaves out dependents' compensation in every year, which moves its two readers:

• Indiana in_unemployment_compensation_deduction (IC 6-3-2-10, in effect from 2021). It deducted a dependent's compensation that is not in the filers' federal AGI, and now does not. Test: in_unemployment_compensation_deduction.yaml Case 3.
• taxsim_v11 (contrib TAXSIM output), which reports the tax-unit total.

Invariants

These hold for every input, net repayments included, and are tested on random households in tests/core/test_unemployment_compensation_exclusion_invariants.py (Hypothesis):

1. Bounds. For each person, taxable ≥ 0, exempt ≤ min(max(0, compensation), 10,200), and taxable + exempt = compensation. A net repayment is neither taxable nor excluded: exempt = compensation. In 2021 nothing is excluded.
2. Cutoff. The head's and spouse's compensation is exempt only when taxable_uc_agi < 150,000, and a dependent's only when their own dependent_taxable_uc_agi is. When it is, the exempt part is exactly min(max(0, compensation), 10,200).
3. Differential. taxable_uc_agi = max(0, adjusted_gross_income − tax_unit_taxable_unemployment_compensation). These are two independent ways of figuring AGI without the compensation, and they agree only if the compensation is left out exactly once.
4. Return boundary. The tax-unit total is the sum over the head and spouse. Zeroing dependents' compensation changes nothing on the filers' return: their taxable compensation, taxable_uc_agi, AGI and the tax-unit total all stay the same.
5. Symmetry. Swapping the head's and spouse's compensation swaps their taxable compensation and leaves the total unchanged.

The property test fails on the old formulas. Hypothesis shrank to wages of 150,000 and compensation of 1, which the old code excluded.

Tests

YAML cases are worked line by line from the 2020 worksheet. On the old formulas, 11 of the 18 cases in the two unemployment test files failed and 7 passed. Two of the 11 also read dependent_taxable_uc_agi, which did not exist yet, but their taxable amounts are wrong on the old code as well. All pass now.

• taxable_unemployment_compensation.yaml Cases 4–14:
  • the Montana repro and the case one dollar under it;
  • the IRS joint example;
  • two spouses over the cap;
  • the dependent repro;
  • a dependent over the cap;
  • a dependent whose own AGI is 150,000;
  • Social Security made taxable by the full compensation (§ 85(c)(2)(A) after § 86; Social Security Benefits Worksheet line 3, file page 31);
  • adjustments;
  • 2021;
  • a net repayment.
• tax_unit_taxable_unemployment_compensation.yaml: the boundary cases are rewritten on the correct reading, with AGI of exactly 150,000, 149,999, and 154,999 (the last was the old file's "full exclusion" case), plus a joint return with a dependent.
• dependent_taxable_uc_agi.yaml: the dependent's own adjustments, and Social Security taxable on the full compensation.
• in_unemployment_compensation_deduction.yaml Cases 2–3. Case 3 fails on the old code.

Local runs: all YAML files above, plus the 2020-affected consumer files (student loan MAGI, IRA phase-out, AR, DE, MA, MT, IN, WI, RI; 73 cases); test_modified_agi_add_backs.py (13 passed); core/test_dependent_above_the_line_deductions_invariants.py (9 passed); code_health/test_rest_python_groups.py (12 passed).

Cost of the new Python test. It runs in the Rest core group. Locally on macOS: 2 tests in 7 s (28 s with the model import), 1.1 GB peak RSS. Each Hypothesis example builds one vectorized simulation, so nothing is retained between tests. These are local measurements, not Linux CI.

Consumers checked

• Federal. irs_gross_income (filers only). taxable_ss_magi, ira_219g_magi, ira_219g_taxable_social_security, student_loan_interest_ald_magi and capped_qualified_tuition_expenses_ald all use the full compensation, as §§ 86(b)(2)(A), 219(g)(3)(A)(ii), 221(b)(2)(C) and the 2020 instructions' line 7 caution require; unchanged. dependent_gross_income overrides with the full compensation; unchanged.

• States, 2020. MA's qualified unemployment deduction (head and spouse, capped), MT's subtraction and UT's homeowner/renter relief now read each person's correct federal amount.

• California, Rhode Island, Wisconsin: land Leave dependents' unemployment compensation off California, Rhode Island and Wisconsin filer returns #10187 (and Leave dependents' income out of Wisconsin homestead credit household income #10191) first. ca_agi_subtractions, ri_additions and wi_unemployment_compensation_subtraction sum every member onto the filers' state return. The independent review of this PR found that once a dependent's 2020 exclusion moves to the dependent's own return, some pooled cases that happened to be right turn wrong. Example: a filer at $160,000 with a dependent's $5,000 of compensation; Rhode Island would add the dependent's excluded $5,000 to the filer's return. Leave dependents' unemployment compensation off California, Rhode Island and Wisconsin filer returns #10187 fixes those three lines. I merged its head 042b072 onto this PR's head 1bf37f4: the merge is clean, its state tests pass, and so do the cases from the reviewer's findings. After Leave dependents' unemployment compensation off California, Rhode Island and Wisconsin filer returns #10187 merges I'll merge main here and add the two 2020 cases that need this PR's dependent exclusion (California: a dependent with $150,000 of dividends; Rhode Island: a filer over the cutoff).

• Other state lines that pool dependents (independent review, round 2). Delaware de_agi_joint, Montana mt_agi_joint, DC dc_taxable_income_joint, Iowa ia_net_income, Utah homeowner/renter and RI-1040H household income, Wisconsin homestead, and SC TANF each read a dependent's person-level compensation onto the filers' return or a household measure, or count taxable rather than received compensation. In 2020 the old pooled allocation made some of them right by accident, and this PR's per-person allocation makes a different set right by accident.

  A 2020 grid of 96 households per state checks each line against its own rule:

  • Households: joint parents with head wages 20k/80k/140k/165k and spouse compensation 0/6k/14k, plus a dependent aged 17 or a 20-year-old student with 4k/12k of compensation and 0/160k of dividends. Each runs with and without the dependent's compensation, on the old formulas and on this PR.
  • Oracle: filer-return lines don't move; household lines that count every member rise by the compensation; Wisconsin homestead (claimant and spouse) doesn't move; SC TANF rises by it / 12 a month.

  State	Line	Right before	Right after	Wrong to right	Right to wrong
  DE	de_agi_joint	24/96	24/96	12	12
  MT	mt_agi_joint	24/96	24/96	12	12
  DC	dc_taxable_income_joint	24/96	24/96	12	12
  IA	ia_net_income	24/96	24/96	12	12
  PR	pr_gross_income_person	48/96	96/96	48	0
  UT	homeowner/renter household income	12/96	12/96	6	6
  RI	ri_property_tax_household_income	24/96	24/96	12	12
  WI	wi_homestead_income	12/96	48/96	42	6
  SC	sc_tanf_gross_unearned_income	36/96	48/96	30	18

  Splitting each cell's absolute error (rounds 3-4 of the independent review) shows where the movement comes from:

  • Ordinary dependents (no dividends): the PR lowers the error by 22% (UT) to 80% for DE, MT, DC, IA, UT and RI-1040H, and no ordinary cell gets worse.
  • The $160,000-dividend dependent: every increase on those six lines comes from this case. Their own § 85(c)(2) AGI is $150,000 or more, so the law makes their compensation taxable on their own return, and the lines pool it onto the filers' return. On the hub's 10% screen sample of the default dataset, dependents with both unemployment compensation and an own-return AGI of $150,000 or more number 0 records and 0 weighted, against about 27,000 weighted dependents with any compensation.
  • WI homestead and SC TANF: their ordinary cells move the other way (30 and 24 cells worse), because WI adds a dependent's exempt compensation and SC counts taxable rather than received compensation.
  • 2021, same grid: DE, MT, DC, IA, UT and RI-1040H fail all 96 cells under both versions.
  • Landing order: after Leave dependents' unemployment compensation off California, Rhode Island and Wisconsin filer returns #10187 (CA, RI, WI state lines) and Leave dependents' income out of Wisconsin homestead credit household income #10191 (WI homestead household = claimant and spouse).
  • Tracked elsewhere: DE/DC in Leave dependents' own items out of DC, Delaware and Ohio subtractions #9939 and Leave dependents' capital income out of state lines on the filers' return #9894; IA in issues AL, AR, MS and NJ add tax unit dependents' income to the filer's state income (Iowa net income too), although federal AGI leaves it out #9814 and Iowa and Mississippi move a dependent's net income onto the head's state return #9887; MT, UT and RI-1040H in follow-up chips; SC TANF (count received compensation) in a follow-up chip; a proposed session for Puerto Rico, whose gross income lists federally taxable compensation that Puerto Rico exempts (the reviewer confirmed this against Schedule IE).
  • Script and results: ~/reviews/uc-exclusion-2020/consumer-grid/.

• 2020 charitable deduction: the $300 deduction for non-itemizers (§ 62(a)(22), Form 1040 line 10b) was taken below the line, so worksheet line 5 was short by it; Take the 2020 charitable deduction for non-itemizers above the line #10186 fixes it.

• Count disability benefits and 2020 excluded unemployment compensation in Montana's elderly homeowner/renter credit income #10163 (stacked, open). At its head 32fc16d it no longer pins the federal allocation: Case 7 supplies taxable_unemployment_compensation as an input. I merged that head into this PR's head eabaefd and its unemployment tests pass: the add-back file (9), disability and unemployment (10), and test_mt_elderly_credit_nontaxable_income.py (13). Merge order doesn't matter.

Hub microsimulation impact (current head)

A real microsimulation on the default dataset, main at the merge base 00ce2281d0 against this PR's head 9fde2d1ec4, for 2024. The runs went one at a time under the shared heavy-job lock. Sums are weighted.

Output	2024 change	2024 records changed (weighted)
unemployment_compensation	+0	0 (0)
taxable_unemployment_compensation	+0	0 (0)
tax_exempt_unemployment_compensation	+0	0 (0)
tax_unit_taxable_unemployment_compensation	-369.26M	27 (90,545)
taxable_uc_agi	+0	0 (0)
adjusted_gross_income	+0	0 (0)
in_unemployment_compensation_deduction	-78	1 (0)
ca_agi_subtractions	+0	0 (0)
ri_additions	+0	0 (0)
wi_unemployment_compensation_subtraction	+0	0 (0)
dependent_taxable_ss_magi	-19	2 (1,221)
state_income_tax	+0	0 (0)
income_tax	+0	0 (0)
household_net_income	+0	0 (0)
household_benefits	+0	0 (0)

Current law, 2024. Only the tax-unit total moves: tax_unit_taxable_unemployment_compensation now leaves out dependents' compensation (27 records). Indiana's deduction moves by $78 on one record (its weight rounds to zero). Dependents' own Social Security modified AGI falls by $19 on 2 records, from netting estate losses. Person-level taxable compensation, AGI, income tax, state income tax and household net income are unchanged. The exclusion's own effect, which outside 2020 is zero, is sized in the counterfactual below.

Comparison files are in fixes/10175-impact/ on the hub host.

Counterfactual: 2020 exclusion rules on 2024 data

The default dataset holds 2024 records only, and core does not carry inputs back to earlier years. On the screen sample, 2020 unemployment_compensation and employment_income both sum to exactly 0. A 2020 pair would therefore show no change whatever the code does.

To size the fix itself, this pair runs 2024 with the 2020 exclusion parameters on both sides: gov.irs.unemployment_compensation.exemption.amount 10,200 and every cutoff 150,000, via IMPACT_REFORM_JSON. It compares main at the merge base 00ce2281d0 with this PR's head 9fde2d1ec4, on the full default dataset under the hub's heavy lock.

These are 2024 incomes, 2024 federal law apart from the exclusion, and 2024 state law on main's state lines, before #10187. They are not an estimate of 2020 itself. Sums are weighted.

Output	Change	Increases	Decreases	Records changed (weighted)
unemployment_compensation	+0	+0	+0	0 (0)
taxable_unemployment_compensation	+254.36M	+592.83M	-338.48M	86 (210,236)
tax_exempt_unemployment_compensation	-254.36M	+338.48M	-592.83M	86 (210,236)
tax_unit_taxable_unemployment_compensation	+254.36M	+592.83M	-338.48M	68 (154,162)
taxable_uc_agi	+0	+0	+0	0 (0)
adjusted_gross_income	+278.22M	+592.83M	-314.61M	60 (147,533)
in_unemployment_compensation_deduction	-4	+0	-4	1 (0)
ca_agi_subtractions	+67.91M	+83.92M	-16.01M	8 (36,153)
ri_additions	-669,649	+0	-669,649	3 (531)
wi_unemployment_compensation_subtraction	+0	+0	+0	0 (0)
dependent_taxable_ss_magi	-19	+0	-38	2 (1,221)
state_income_tax	+19.88M	+20.84M	-961,628	38 (80,744)
income_tax	+96.44M	+134.19M	-37.75M	58 (147,533)
household_net_income	-122.40M	+38.72M	-161.11M	60 (147,533)
household_benefits	+0	+0	+0	0 (0)

• Why taxable compensation rises. From the formulas, a filer's taxable compensation can rise only by losing the exclusion. That happens where modified AGI is $150,000 or more and the old code subtracted the compensation a second time to get under the cutoff.
• Why it falls. For the same eligibility, the old per-person exclusion was own × 10,200 / unit total, which is never more than the new min(own, 10,200). It falls where both spouses received compensation, or where a dependent's compensation had shared the filers' cap.
• Why AGI moves less. AGI's decreases ($314.6M) are smaller than taxable compensation's ($338.5M). The difference is dependents' compensation, which was never in the filers' AGI.
• California and Rhode Island. California's subtraction tracks the filers' taxable compensation. Rhode Island's add-back of excluded compensation falls by $0.67M, because, on main's pooled line, dependents' exempt amounts change.

Files: ~/reviews/uc-exclusion-2020/impact/rules2020-on-2024/ on the hub host; reform JSON at ~/reviews/uc-exclusion-2020/impact/reform_2020_rules_2024.json.

axiom: TheAxiomFoundation/rulespec-us#1719 queued. us:statutes/26/85 already encodes § 85(c)(1) correctly: per-recipient $10,200 cap, joint return summed per spouse, strict <, with tests temporary_2020_exclusion_applies_per_spouse_on_joint_return and temporary_2020_exclusion_denied_at_threshold. But the § 85(c)(2) AGI and the recipient relation are free inputs. #1719 queues encoding both, with companion tests from the same worksheet cases.

🤖 Generated with Claude Code

[MaxGhenis]

Hand-off to the US + core hub. Session bebaaf35, head 9fde2d1ec4.

• Independent review: APPROVE at 9fde2d1ec4. Round 4 was served by Opus 5.5 because the Codex lanes were closed; rounds 1–3 were GPT-6.1 Sol. Reports: ~/reviews/uc-exclusion-2020/review_out.md, review2_out.md, review3_out.md, review4_out.md; author responses response_round2.md and response_round3.md.
  • Every round-1 and round-2 finding is fixed, accepted as out of scope, or withdrawn with evidence:
    • Fixed: the § 461(l) limit on the dependent's own return, non-negative taxable UC, and the citations.
    • Out of scope: the WI homestead, SC TANF and PR consumers.
    • Withdrawn: the DE/MT/DC/IA/UT/RI-1040H pooling findings. The per-cell error split and the screen-sample count show every worse case needs a dependent with both UC and an own AGI of $150k or more, and there are 0 such dependents.
  • On a merged tree of this head + Leave dependents' unemployment compensation off California, Rhode Island and Wisconsin filer returns #10187 208a87f7a9 + Leave dependents' income out of Wisconsin homestead credit household income #10191 509e5c9d71, the reviewer found both merges clean, 82/82 YAML cases passing in the 13 touched files, and 8/8 invariant tests passing.
• CI: all 26 checks pass at 9fde2d1ec4 (gh pr checks exits 0). Mergeable.
• Landing order: land after Leave dependents' unemployment compensation off California, Rhode Island and Wisconsin filer returns #10187 (CA/RI/WI state lines) and Leave dependents' income out of Wisconsin homestead credit household income #10191 (WI homestead household). This PR gives each dependent their own 2020 exclusion. Those two PRs keep a dependent's amounts off the filers' state lines. The branch does not need main merged in (no conflicts; branch protection is not strict).
• Impact: both full-file pairs ran at this head; tables are in the body. Current law 2024 moves no tax. The 2020 rules applied to 2024 data move AGI by +$278M and income tax by +$96M. A literal 2020 run would read all-zero inputs on the 2024 dataset.
• Axiom: [pe-parity] Encode IRC §85(c)(2) AGI for the 2020 unemployment exclusion and limit recipients to taxpayer and spouse (PE-US #10175) TheAxiomFoundation/rulespec-us#1719 is queued (pe-parity) for § 85(c)(2) AGI and the recipient relation.
• No hold, no decision. It is a bug fix to the statute and the IRS worksheet.
• Follow-ups, not blocking:
  • session chips for DE/DC/IA/PA/MT/UT pooling of dependents' UC and for SC TANF counting taxable UC;
  • a proposed session for Puerto Rico gross income including federally taxable UC;
  • Take the 2020 charitable deduction for non-itemizers above the line #10186 (2020 $300 charitable deduction above the line);
  • Leave dependents' own items out of DC, Delaware and Ohio subtractions #9939/Leave dependents' capital income out of state lines on the filers' return #9894 (DE/DC);
  • issues AL, AR, MS and NJ add tax unit dependents' income to the filer's state income (Iowa net income too), although federal AGI leaves it out #9814/Iowa and Mississippi move a dependent's net income onto the head's state return #9887 (IA) and South Carolina TANF counts taxable UC and taxable Social Security instead of gross amounts #9677 (SC TANF).