[MaxGhenis]

Models three South Carolina TANF rules that #10184 left out. Fixes the South Carolina part of #10497; Michigan FIP's half stays open there.

Rules

1. Dependent children's earned income is excluded (S.C. Code Regs. 114-1140(N)(1); TANF Policy Manual Vol. 67 §7.5, PDF p.105; §8.16 B, p.134). 114-1140(N)(1) excludes "the earned income of dependent children" in "the FI gross income limit test and the need and benefit determination". §7.5's first row says "Earned income of a dependent child(ren) is excluded". A minor parent's earnings count "if the minor parent is the head-of-household" and are excluded "if the minor parent is included as a dependent in the TANF BG". The exclusion sits in sc_tanf_gross_earned_income, so the gross test, the net test and the benefit all see it. That variable's only readers are sc_tanf_gross_income and sc_tanf_countable_earned_income_person, and neither needs the child's earnings. A dependent child's unearned income still counts.

Mapping (new sc_tanf_dependent_child). 114-1130(E) (p.204) sets the age test: a dependent child "must be under eighteen years of age or under nineteen years of age and a full-time student in a secondary school". The model reuses gov.hhs.tanf.cash.eligibility.age_limit: under 18, or under 19 when is_in_secondary_school. A person who meets it and is the head or spouse of their tax unit (is_tax_unit_head_or_spouse) is treated as a head-of-household minor parent, so their earnings count; everyone else who meets it is a dependent child, apart from the no-adult rule below. Pregnancy plays no part: is_person_demographic_tanf_eligible also returns true for any pregnant person, which would wrongly exclude a pregnant adult's wages. Policy Manual §4.10 (p.75) extends the age to a full-time student "in a primary or secondary school or equivalent", who "may receive benefits until age 19". The model follows the regulation's "secondary school", so it treats an 18-year-old in primary school as a non-student. It doesn't test "unmarried" (§4.10). An 18-year-old in secondary school meets the age test but is an adult for role inference, so with no dependent input one who lives with a single parent is inferred to be the tax unit's spouse, and their earnings count.

In a group with an adult, a minor, married or not, counts as a parent only when the household input names them the tax unit's head or spouse, because PE-US infers tax-unit roles only for adults. In a group with no adult where no member holds a head or spouse role, the oldest member heads the household, and members who tie for oldest all head it. Where a member does hold one of those roles, the inputs decide. A 17-year-old input as a younger head's tax dependent is a dependent child, and 114-1140(N)(1) excludes that 17-year-old's earnings (sc_tanf_dependent_child.yaml Case 9, integration.yaml Case 31). The default dataset supplies every person's tax-unit role, and all 29 people in its 28 SPM units with no adult hold the head role, so these limits change household calculations and not the microsimulation.

2. Child support paid outside the group comes off the payer's own income (S.C. Code Regs. 114-1140(L); §8.11, p.131; §8.16 D and F, p.134). 114-1140(L) says support "made by a budget group member to legal dependents who are not budget group members may be deducted from the income of the payor". 114-1130(B)(3) reduces a stepparent's own income the same way. New sc_tanf_child_support_deduction takes each member's child_support_expense, read monthly, from that member's own income: earned income after the §8.12 disregards, then unearned income, in the order of §8.16 C, D and F, never below zero. It sums those amounts over the group.

• Any member can pay, a dependent child included, because the regulation says "a budget group member". A dependent child's earnings are excluded, so the child's support comes off the child's unearned income (integration.yaml Case 33).
• §7.7 excludes an SSI recipient's income, so an SSI recipient's support is never deducted (Case 32).
• A payer with no counted income gets no deduction (Case 22), which covers §8.11(2)'s group with no income.
• sc_tanf_countable_income subtracts the deduction, which serves the §8.16 H net test and the benefit. sc_tanf_gross_income doesn't read it, so the gross test stays on gross income.

Countable income keeps its cents; §8.1 (p.121) says "Do not round at any step of the TANF calculation". As on main, a self-employment loss in a month of the 50% disregard can take countable income below zero (Case 35). The deduction never takes a payer's own income below zero, so another member's loss doesn't cancel a payer's deduction (Case 36). The manual reads as pooling: §8.16 D subtracts support "from the remaining earned income of any adult", F takes support "paid by any adult" from the unearned income, and §8.11 deducts it "as income to the BG". The regulation governs, and source_conflicts/candidates.yaml records the difference as us-sc:tanf-child-support-deduction-payer.

3. Whole dollars (S.C. Code Regs. 114-1140(A)(4); §8.15 table notes, p.133). 114-1140(A)(4) rounds the benefit "down to the next lower whole dollar" when "the standard of need or the payment amount is not a whole dollar". The table notes say "multiply the Need Standard by .4046 and drop the cents" for the payment standard. The changes:

• the need standard becomes ⌊50% × monthly FPG⌋;
• the benefit becomes ⌊min((need − countable) × 0.4046, need × 0.4046)⌋, floored before the $10 minimum test.

The table rounds the monthly FPG before halving it ($1,762 → $881 for two). ⌊⌊x⌋/2⌋ = ⌊x/2⌋ for x ≥ 0, so flooring once gives the same result. Countable income keeps its cents.

The gross income limit keeps its cents too. 114-1150(A) (p.217) says gross income "must not exceed 185 percent of the need standard by family size", and 114-1140(A)(4) rounds only the need standard and the payment. The §8.15 note says "multiply the Need Standard by 1.85 and drop the cents", which fails groups up to a dollar below the regulation's limit. The model follows the regulation, so the limit is 1.85 times the floored need standard: $1,629.85 for two in 2026, where the table prints $1,629. source_conflicts/candidates.yaml records the difference as us-sc:tanf-gross-income-limit-cents.

Above eight persons, the §8.15 note adds $229 for each person, while 114-1140(A)(2) sets the standard at "fifty percent of the current year's" poverty guidelines. The two agree through 12 persons and differ from 13: $3,402 under the regulation and $3,401 under the note for 13 persons in 2026. The model follows the regulation, and source_conflicts/candidates.yaml records the difference as us-sc:tanf-need-standard-above-eight.

Out of scope. §8.10 A.2 (p.131) subtracts child support paid when a parent's income is deemed to a minor parent. The model doesn't deem income to minor parents, so that rule is left as is. Michigan FIP's rounding is the other half of #10497.

Expected values

2025 FPG ($15,650 + $5,500 per added person), except the two #10497 rows, which use the 2024 guideline ($15,060 + $5,380) in effect in January 2025. Every household is in South Carolina.

Case	Main	This PR
§8.15 need, sizes 1–8 (2026-01)	652.08 … 2,256.25	652, 881, 1,110, 1,339, 1,568, 1,797, 2,027, 2,256
§8.15 gross limit, sizes 1–8	1.85 × the unfloored need (1,630.31 for 2)	1,206.20, 1,629.85, 2,053.50, 2,477.15, 2,900.80, 3,324.45, 3,749.95, 4,173.60
§8.15 payment standard, sizes 1–8	unfloored	263, 356, 449, 541, 634, 727, 820, 912
Size 9 need	2,485.42	2,485 ($229 above size 8)
Parent + 2 children, no income, 2026-01	449.27	449
Same, child 16 earns $500/mo, 2026-01	348.12	449
Same, parent earns $600/mo and pays $2,400/yr, 2026-05	246.97 (countable 500)	327 (countable 300; gross stays 600)
Two adults + child: A pays $2,400/yr and has no income, B earns $600/mo, 2026-05 (Case 22)	246.97 (countable 500)	246 (countable 500)
Parent earns $250/mo, $100/mo alimony, pays $3,600/yr, 2026-05	348.12 (countable 250)	449 (countable 0)
Parent earns $600/mo and pays $2,400/yr, 2026-01	327.89 (countable 300)	408 (countable 100)
Minor parent, 17, earns $500/mo, lives with a baby and no adult, no role inputs, 2026-01	255.40 (countable 250)	255 (countable 250)
Spouse 20; spouse 17 input as tax-unit spouse, earns $500/mo; baby; 2026-01	348.12 (countable 250)	347 (countable 250)
Minor 16 input as tax-unit head; 17-year-old input as that head's dependent, earns $500/mo; baby; 2026-01 (Case 31)	348.12 (countable 250)	449 (countable 0)
SSI parent pays $2,400/yr; other parent earns $600/mo; child; 2026-05 (Case 32)	154.25 (countable 500)	154 (countable 500)
Parent; dependent child 17 with $200/mo of survivor benefits pays $1,200/yr; child; 2026-05 (Case 33)	368.35 (countable 200)	408 (countable 100)
A earns $600/mo and pays $2,400/yr; B has a $14,400/yr self-employment loss; child; 2026-01 (Case 36)	449.27 (countable −300)	449 (countable −500)
Parent + child, wages $1,629.50/mo, 2026-01 (Case 37)	26.91	26 (passes the 1,629.85 limit)
#10497: size 2, Jan 2025, countable $820	12.81	12
#10497: size 2, Jan 2025, countable $826.80	10.06	0

All values were computed by hand, and each YAML case shows its arithmetic. Every fractional expectation already in the SC files was recomputed the same way (sc_tanf.yaml, integration.yaml, sc_tanf_need_standard.yaml, sc_tanf_gross_income_eligible.yaml, sc_tanf_countable_income_eligible.yaml, and the need standard in test_sc_tanf_unearned_income_invariants.py). The main column of each row that names a case comes from a run at origin/main ba9bfdea75; the other rows' main values come from earlier rounds' runs.

Invariants (Hypothesis, test_sc_tanf_child_earnings_support_invariants.py)

For every generated South Carolina benefit group:

• (a) sc_tanf is a whole number of dollars, and is 0 or at least $10.
• (b) With no self-employment loss, countable income never falls below 0. It never rises when child support paid rises, and $1 more support lowers it by at most $1.
• (c) Changing a dependent child's earned income never changes sc_tanf_gross_income, sc_tanf_countable_income or sc_tanf.
• (d) sc_tanf never rises when an adult's wages rise.
• (e) Child support paid never changes sc_tanf_gross_income.
• (f) With no dependent-child earnings and no child support paid, sc_tanf_countable_income equals main's formula to the cent. The test freezes that formula as a reference function and checks it on generated groups and on 200 seeded South Carolina households in each of January and May.

Generated groups have no self-employment income, nobody receives SSI, and only adults pay support, so earned income is never negative. Every generated group has an adult head, so YAML cases alone pin the no-adult rule, a dependent child's or SSI recipient's support (Cases 32 and 33), and a self-employment loss (Cases 35 and 36).

Mutants (one process per test file, source restored after each)

Every row ran this round in a scratch worktree. Rows 1–12 are earlier rounds' mutants, rerun at the final head e4e6861e6a against integration.yaml, sc_tanf_dependent_child.yaml, sc_tanf.yaml, sc_tanf_need_standard.yaml, sc_tanf_gross_income_eligible.yaml, sc_tanf_countable_income_eligible.yaml, sc_tanf_gross_earned_income.yaml and the property file. Rows 2 and 4 are rewritten for the per-payer deduction. Rows 13–15 ran at 675ca2d115, whose code and YAML match the final head, against all 15 SC TANF YAML files and both property files. Rows 16–21 ran there against integration.yaml and the property file, plus both gross income files for rows 19–21. Row 20's property run and row 15's case names are from the final head. git status was clean after each.

#	Mutant	Result	Killed by
1	Drop the dependent-child exclusion	killed	integration.yaml Cases 19, 20, 30, 31; sc_tanf_dependent_child.yaml Cases 4, 8, 9; property (c)
2	Take each payer's support from that payer's gross earnings (floored at zero), apply the §8.12 disregard, then take the rest from unearned income	killed	integration.yaml Cases 21, 23, 25, 35, 36. Case 25 (January) is the ordering case; Case 21 fails on its asserted deduction, since this mutant moves the support into earned income.
3	Let child support reduce sc_tanf_gross_income	killed	integration.yaml Cases 21, 23, 25 (gross income); property (e)
4	Limit each payer's deduction to that payer's earned income	killed	integration.yaml Cases 23, 33
5	Drop the need standard floor	killed	sc_tanf_need_standard.yaml 11 of 12; integration.yaml Cases 1, 2, 5, 6, 15, 18, 26–29, 32; sc_tanf.yaml Case 8; sc_tanf_gross_income_eligible.yaml Case 5 and the eight "one cent above" cases; sc_tanf_countable_income_eligible.yaml Case 4
6	Apply the $10 test before flooring the benefit	survived (equivalent)	⌊x⌋ ≥ 10 exactly when x ≥ 10, so the order can't change a payment
7	Treat a pregnant adult as a dependent child	killed	integration.yaml Case 24; sc_tanf_dependent_child.yaml Case 3; both (f) differentials
8	Drop the no-adult clause (heads_household = head_or_spouse)	killed	sc_tanf_dependent_child.yaml Cases 5, 6, 7, 8; integration.yaml Cases 26, 28, 29, 30
9	Every member of a group with no adult heads it (oldest replaced with True)	killed	sc_tanf_dependent_child.yaml Cases 5, 6, 7, 8; integration.yaml Cases 26, 28, 29, 30. Case 2 no longer kills it, because an input head now turns the rule off.
10	Drop the head test (return meets_age_test)	killed	sc_tanf_dependent_child.yaml Cases 2, 5–9; integration.yaml Cases 26–31
11	Pick one oldest member by rank (oldest = person.get_rank(person.spm_unit, -age) == 0)	killed	sc_tanf_dependent_child.yaml Cases 6, 7; integration.yaml Cases 28, 29
12	Make every member within a year of the oldest a head (oldest = age >= person.spm_unit.max(age) - 1)	killed	sc_tanf_dependent_child.yaml Case 8; integration.yaml Case 30
13	Drop the no_head term (head_or_spouse | (no_adult & oldest)), the reading before this round	killed	sc_tanf_dependent_child.yaml Case 9; integration.yaml Case 31
14	Pool the deduction across the group: min_(sum(support), max_(sum(income), 0))	killed	integration.yaml Cases 22, 32, 36
15	Restore the floor on the gross income limit	killed	sc_tanf_gross_income_eligible.yaml Case 4 and the eight "one cent below" cases; integration.yaml Case 37
16	Restore the SSI payer condition	survived (equivalent)	An SSI recipient's counted income is already 0, so the cap gives 0 with or without it. Case 32 pins the result.
17	Restore the dependent-child payer condition	killed	integration.yaml Case 33
18	Drop the max_(child_support_expense, 0) clamp	killed	integration.yaml Case 34
19	Floor countable income at 0 instead of the members' negative incomes	killed	integration.yaml Cases 35, 36
20	Drop the countable income floor (income - deduction)	killed	property (b), on the fixed January example. At 675ca2d115, before that example, random draws missed it and it survived.
21	Restore the old where(deduction > 0, max_(income - deduction, 0), income)	killed	integration.yaml Case 36

The property tests killed rows 1, 3, 7 and 20 only. Every generated group has an adult head, only adults pay support, and no member has a loss. Row 13 was this body's mutant 13 before this round: the code now takes that reading, so the reverse mutant is the check.

The property tests found one real bug in round 1: countable income came out a float32 rounding error below zero when a deduction used up a payer's income. The stored float32 deduction can exceed the float64 income sum by about 3e-5. sc_tanf_countable_income floors the result at the sum of the members' negative incomes. That sum is zero unless a member has a self-employment or farm loss. Generated groups have no loss and no negative income, so the floor leaves (f) exact.

Tests

Each YAML file ran in its own process at 675ca2d115, whose code and YAML match the final head; the final head adds one example to the property file, which ran there. Every SC TANF YAML file passes, 145 cases in all:

File	Cases
integration.yaml	37
sc_tanf_gross_income_eligible.yaml	21
sc_tanf_gross_unearned_income.yaml	18
sc_tanf.yaml	16
sc_tanf_need_standard.yaml	12
sc_tanf_dependent_child.yaml	9
sc_tanf_countable_earned_income_person.yaml	4
sc_tanf_countable_income_eligible.yaml	4
sc_tanf_eligible.yaml	4
sc_tanf_income_eligible.yaml	4
sc_tanf_resources_eligible.yaml	4
sc_tanf_ssi_recipient.yaml	4
sc_tanf_benefit_group_size.yaml	3
sc_tanf_countable_resources.yaml	3
sc_tanf_gross_earned_income.yaml	2

The no-adult cases:

• sc_tanf_dependent_child.yaml Case 5 and integration.yaml Case 26: a minor parent who lives with no adult and has no role inputs.
• Cases 6 and 7, and integration Cases 28 and 29: two 17-year-olds, one earning, and a baby, with no role inputs, listed in two orders.
• Case 8 and integration Case 30: a 17-year-old, an earning 16-year-old and a baby, with no role inputs. No input says which minor is the parent, so they pin the inference rule.
• Case 9 and integration Case 31: the same three people with the 16-year-old input as the tax unit's head. The 17-year-old is a dependent child, and the payment is 449.
• integration Case 27: a married minor input as the tax unit's spouse.

The support cases: integration Cases 21, 23 and 25 have one payer who holds the group's income. Case 22 has a payer with no income, Case 32 an SSI recipient who pays, and Case 33 a dependent child who pays from survivor benefits. Case 34 has a negative support input, and Cases 35 and 36 a self-employment loss without and with a payer. Case 37 and the gross test file's boundary cases pin the limit's cents.

Python:

• test_sc_tanf_child_earnings_support_invariants.py passes 4 at the final head. Under /usr/bin/time -v it took 43.5 s wall clock including startup (7.7 s of tests) and peaked at 1,054,820 KB RSS. Before this round, at 96bd29a40f, it took 43.5 s and peaked at 1,052,088 KB. Both are local runs on a 4 vCPU, 16 GB Linux VM with Python 3.14, not CI measurements. It builds simulations of small groups only, with no dataset. It sits outside core/ and policy/ and isn't a listed heavy module, so it runs in the Makefile's remaining Python group, which the Makefile records at 14.4 GB and 22:18 in CI run 38141803955.
• test_sc_tanf_unearned_income_invariants.py passes 2.
• code_health/test_source_conflicts.py passes 103.
• test_known_dead_reference_urls.py passes 40, test_reference_urls.py 32 and test_no_concatenated_reference_urls.py 48.

ruff format --check and ruff check pass on the PR's nine Python files with the locked ruff 0.15.5. No partner test reads SC TANF, and none was edited; nothing under .github/ changed. No microsimulation was run.

CI: Pull request run 42641205118 passed all 36 jobs at e4e6861e6a; none was skipped or cancelled.

Overlap

Trial merges of this head with git merge-tree --write-tree, against the current head of every open PR that touches source_conflicts/candidates.yaml, an SC TANF file or is_tax_unit_head.py (35 PRs, listed on 2026-11-09):

• None conflicts with this PR in a file where it doesn't already conflict with main.
• Six already conflict with main in candidates.yaml: Model the Medicare Part D low-income subsidy as a CBO means-tested transfer #10127, Fix Arkansas tax indexing and Kentucky's 2021 dependent care credit #10324, Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527, Apply each state's rule for a Medicaid applicant whose spouse is not aged, blind or disabled #10640, Apply the New York itemized deduction adjustment to IT-196 line 45 #10658 and Round NYC EIC worksheet reductions #10664. Against this head each has the same number of conflict hunks, and no hunk holds a us-sc:tanf-* id.
• Include farm and partnership earnings in SSI and correct loss netting #10532 conflicts with main, and with this PR, only in ca_state_supplement.yaml. Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 (draft) also conflicts in Makefile, me_eitc.py, nj_eitc.py and nj_eitc.yaml, the same files against main.
• Keep an inferred tax unit spouse in the head's marital unit #10555 changes how is_tax_unit_head.py picks a head for adult couples, and merges cleanly. Its couple candidates must already be adult candidates, so the no-adult rule here is unaffected.
• Round summed resources to whole cents before exact-limit asset tests #10639 (SC TANF resources) merges cleanly.
• Add South Carolina LIHEAP regular heating assistance #10616 (SC LIHEAP) merged into main as 5783608cd1. This round's merge of main kept its six us-sc:liheap-* entries ahead of this PR's three us-sc:tanf-* entries.

impact: pending a comparison on the default dataset for 2026, main against this branch.
axiom: TheAxiomFoundation/rulespec-us#1728 queued

🤖 Generated with Claude Code

[MaxGhenis]

Ready for hub review

Head: 600d93e70acd8c9b6adbeb13d5100a9d98a1d188. The PR is a draft, branched from origin/main 97b6d5135c, which contains #10184 (563d66c04d, confirmed with git merge-base --is-ancestor).

Rules and where they are

• Rule 1. sc_tanf_dependent_child.py (new) and income/sc_tanf_gross_earned_income.py:15-25. The exclusion sits in gross earned income, whose only readers are sc_tanf_gross_income and sc_tanf_countable_earned_income_person; neither must keep a child's earnings. Dependent-child definition: not found. The manual PDF was unreachable from this VM (dss.sc.gov returns 403 at the proxy), so I couldn't quote it. The model uses gov.hhs.tanf.cash.eligibility.age_limit (under 18, or under 19 with is_in_secondary_school). A head or spouse of the tax unit who meets that test counts as a head-of-household minor parent. Pregnancy isn't read. The age mapping is unverified against the manual.
• Rule 2. income/sc_tanf_child_support_deduction.py (new) returns min(S, max(E + U, 0)); income/sc_tanf_countable_income.py subtracts it. Reader check: sc_tanf_countable_income is read by sc_tanf_countable_income_eligible (§8.16 H) and sc_tanf. sc_tanf_gross_income (the §8.15 test) adds only gross earned and gross unearned income, so the deduction never reaches it.
• Rule 3. sc_tanf_need_standard.py (floor, with the ⌊⌊x⌋/2⌋ = ⌊x/2⌋ comment), eligibility/sc_tanf_gross_income_eligible.py (⌊1.85 × need⌋), sc_tanf.py (⌊capped benefit⌋ before the $10 test). Countable income keeps its cents.

Expected values (hand-computed; head = this PR, run; main = origin/main, run)

Case	Hub	Mine	Head	Main
§8.15 need, sizes 1–8	652 … 2,256	same	same (8 YAML cases)	652.08 … 2,256.25
§8.15 gross limit, sizes 1–8	1,206 … 4,173	same	same (pass at the limit, fail one cent above)	unfloored
§8.15 payment standard, sizes 1–8	263 … 912	same	same	unfloored
Size 9 need	2,485	2,485	2,485	2,485.42
Parent + 2 children, no income, 2026-01	449	449	449	449.27
Child 16 earns $500/mo, 2026-01	449	449	449	348.12
Parent earns $600/mo, pays $2,400/yr, 2026-05	327	327	327 (gross 600)	246.97
Two adults (A pays, B earns $600), 2026-05	327	327	327	246.97
Support above income, 2026-05	449	449	449 (countable 0)	348.12
Same parent, 2026-01 (added: disregard before the deduction)	—	408	408 (countable 100)	327.89
#10497 size 2, Jan 2025, countable 820	12	12	12	12.81
#10497 size 2, Jan 2025, countable 826.80	0	0	0	10.06

I found no disagreement with the hub's values. Every existing fractional SC expectation was recomputed by hand, and each file shows its arithmetic.

Tests

All 15 SC TANF YAML files pass (128 cases); test_sc_tanf_unearned_income_invariants.py passes 2 and test_sc_tanf_child_earnings_support_invariants.py passes 4. The new property file takes 49 s locally with startup (11 s of tests) and uses no dataset. Invariants (a)–(f) are stated in the body and property-tested with Hypothesis on generated groups. The (f) differential checks main's formula, frozen as a reference function, on generated groups and on 200 seeded SC households in each of January and May. No partner file reads SC TANF; none touched. No microsimulation.

Mutants

1, 2, 3, 4, 5 and 7 are killed; the body's table lists which tests kill each. 6 survives because it's equivalent: ⌊x⌋ ≥ 10 exactly when x ≥ 10, so flooring before or after the $10 test can't change a payment. Mutant 2 as first written (capping the deduction by gross earnings) was too weak, so I reran it as "deduct support from gross earnings before the disregard"; Cases 21, 22, 23 and 25 kill it.

CI

Not run: the PR is a draft.

Judgment calls

• Negative earned income. The deduction is min(S, max(E + U, 0)), and countable income floors at zero only when a deduction applies. The brief's max(E − S, 0) + max(U − max(S − E, 0), 0) equals that whenever E ≥ 0. If a business loss makes E negative, main's countable income can already be negative. I left that unchanged, so (f) holds exactly, and the generated cases keep earned income nonnegative.
• Rounding bug the properties found. When the deduction equalled the group's income, countable income came out a float32 rounding error below zero. The zero floor above fixes it.
• May YAML cases use period: 2026 with May-keyed outputs. With a month period, the YAML runner stores yearly inputs for the year that starts that month, which May 2026 can't see.
• Minor-parent mapping follows the brief: tax-unit head or spouse stands in for head of household. I found no better input.
• Body wording. The body's §7.5, 114-1140(A)(4) and §8.15 quotes are the brief's, not fetched here.

Draft comment for TheAxiomFoundation/rulespec-us#1728 (for the hub to post)

#10581 adds these South Carolina TANF provisions. Each needs a RuleSpec module under us-sc/:

• SC DSS TANF Policy Manual Vol. 67 §7.5 (PDF p.105): "Earned income of a dependent child(ren) is excluded". A minor parent's earnings count "if the minor parent is the head-of-household".
• §8.11 (p.131): child support paid outside the home by a benefit group member is deducted; a group with no income gets none (§8.11(2)).
• §8.16 B–G (p.134): the budget order. Dependent-child earnings are excluded (B); the §8.12 disregards come before the support deduction (D, F), and support comes off earned income first, then unearned.
• S.C. Code Regs. 114-1140(A)(4) (Chapter 114, p.210): the standard of need and the payment are "rounded down to the next lower whole dollar".
• §8.15 table notes (p.133): "multiply the Need Standard by 1.85 and drop the cents"; "multiply the Need Standard by .4046 and drop the cents".

Companion cases, computed by hand from those sources (2025 FPG, South Carolina):

1. Benefit group of three, no income, January 2026: need ⌊26,650 / 12 × 0.5⌋ = 1,110; payment ⌊1,110 × 0.4046⌋ = $449.
2. Same group, May 2026. The parent earns $600 a month and pays $2,400 a year in child support: E = 600 − 100 = 500, S = 200, countable 300, payment ⌊810 × 0.4046⌋ = $327. Gross income stays $600.
3. Group of two, January 2025 (need ⌊851.67⌋ = 851), countable income $826.80: ⌊24.20 × 0.4046⌋ = $9, below the $10 minimum, so $0.

Proposed follow-up

• SC TANF: verify the dependent-child definition against Policy Manual Vol. 67. This PR maps it to the federal TANF minor-child ages and the tax-unit head or spouse. Check the manual's own age, school-attendance and minor-parent wording, and §8.10 A.2 deeming to minor parents.

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (aae7f98) to head (e4e6861).
⚠️ Report is 89 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff            @@
##              main    #10581   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            1         7    +6     
  Lines           37       118   +81     
  Branches         0         1    +1     
=========================================
+ Hits            37       118   +81     

Flag	Coverage Δ
unittests	100.00% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[MaxGhenis]

Ready for hub review
VERDICT: REQUEST_CHANGES
cdf464a
METHODOLOGY_CHOICE: yes

Independent review of the whole diff at cdf464ad90 (three commits on main 97b6d5135c). The head hasn't moved. The PR merges cleanly into current main (f9dcb5e7). CI at this head passed every job, including Household API Partners. I found no earlier **Hub review in progress** comment, so this review starts from the beginning.

METHODOLOGY_CHOICE is yes because the PR picks readings where the sources leave room: the tax-unit head or spouse stands in for "head-of-household", one adult's support comes off the whole group's income, and the regulation governs over the manual above eight persons.

Findings

P1. A minor parent who lives with no adult is treated as a dependent child, so her earnings drop out

policyengine_us/variables/gov/states/sc/tanf/sc_tanf_dependent_child.py:33-34

The mapping relies on is_tax_unit_head_or_spouse. When tax-unit roles aren't input, is_tax_unit_head and is_tax_unit_spouse pick only from adults (head_or_spouse_candidates uses ~is_child, that is age 18 or over). A 17-year-old mother living alone with her baby heads no tax unit, so she meets the age test, is not a head or spouse, and is classed as a dependent child. Section 7.5, as the body quotes it, counts a minor parent's earnings "if the minor parent is the head-of-household". In this household nobody else could be the head.

sc_tanf_dependent_child.yaml Case 2 hides this, because it inputs is_tax_unit_head: true for the minor parent. Households entered through the API or the YAML runner don't usually set that input.

Witness, January 2026, mother 17 with $6,000 a year of wages and a baby, no role inputs:

head:  is_tax_unit_head=[False, False]; sc_tanf_dependent_child=[True, True]; sc_tanf_gross_earned_income=[0, 0]; sc_tanf=356
main:  sc_tanf_countable_income=250; sc_tanf=255.40

By hand: $500 a month of wages, a 50% January disregard leaves 250, and floor((881 − 250) × 0.4046 = 255.30) = $255. The head pays $356, the full payment standard for two. Main pays $255.40, which is right apart from rounding.

Fix. I tested this replacement for lines 33-34 in a scratch tree. It gives sc_tanf_dependent_child [False, True] and sc_tanf 255 for the witness, and the married-minor and dependent-minor-parent cases don't change:

        head_or_spouse = person("is_tax_unit_head_or_spouse", year)
        # Tax-unit roles are inferred only for adults, so a minor parent who
        # lives with no adult heads no tax unit unless roles are input. The
        # oldest member of a group with no adult heads the household.
        no_adult = ~person.spm_unit.any(~person("is_child", year))
        oldest = person.get_rank(person.spm_unit, -age) == 0
        heads_household = head_or_spouse | (no_adult & oldest)
        return meets_age_test & ~heads_household

Add a case to sc_tanf_dependent_child.yaml, with no role inputs, that expects [false, true] and gross earned income [500, 0]. Also add an integration.yaml case for the same household that expects sc_tanf: 255.

P2. A married minor whose spouse is an adult is a dependent child unless roles are input, which contradicts a claim in the body and in the hub's mapping ruling

sc_tanf_dependent_child.py:33-34; PR body, Rule 1 mapping paragraph.

The body says: "A married minor is almost always the head or spouse of their own tax unit, which the mapping already treats as a parent." With inferred roles that isn't true. is_tax_unit_spouse also draws only from adults, so a 17-year-old married to a 20-year-old is neither head nor spouse. Declaring the couple as a marital unit doesn't change that.

Witness, January 2026, spouse 20, wife 17 with $6,000 a year of wages, baby:

no role inputs:           is_tax_unit_spouse=[F, F, F]; sc_tanf_dependent_child=[F, T, T]; sc_tanf=449   (main 348.12)
marital unit [dad, mom]:  is_tax_unit_spouse=[F, F, F]; sc_tanf_dependent_child=[F, T, T]; sc_tanf=449

Section 4.10 requires a dependent child to be "unmarried", as the body quotes it, so her wages should count: floor((1,110 − 250) × 0.4046 = 347.96) = $347. I tried reading marriage from marital_unit.nb_persons(), but that isn't safe: when no marital units are input, everyone shares one unit, and the baby came out "married". So I don't propose a code change here. Instead:

• Replace that body sentence with: "The model doesn't test 'unmarried'. A married minor counts as a parent only when they are input as the tax unit's head or spouse, because tax-unit roles are inferred only for adults."
• Add an explicit-role YAML case (is_tax_unit_spouse: true on the 17-year-old) that expects her earnings to count, so the disclosed path is pinned.

Whether the default microdata supplies tax_unit_role_input for minors is under "What I couldn't check". If it does, both this finding and P1 are limited to household-level use, which is still where API users get their numbers.

P2. The body's status lines are false at this head

PR body, "Status" line and the last sentence of "Tests".

The body says "Status: draft … CI has not run (drafts skip it)" and "CI has not run, because this is a draft." The API reports draft: false. Workflow run 41463346352 at cdf464ad90 passed every job: the Baseline, Contrib and Python shards, Microsimulation, Household API Partners, PR validation, Package and compatibility, and Quick Feedback. Replace both with: "CI passed at cdf464ad90."

P3. The hub's population expectation that "the floors lower each payment by under $1" is wrong at two margins

This isn't a code defect. It matters for reading the impact pair.

1. Two floors stack. Flooring the need standard costs up to 0.4046 × (its cents), and flooring the payment costs up to $1 more. Size 6, January 2026, countable $64.43 (alimony): head sc_tanf = 700, main 701.37, so the loss is $1.37. By hand: floor((1,797 − 64.43) × 0.4046 = 701.00) = 700, against (1,797.92 − 64.43) × 0.4046 = 701.37.
2. The gross-limit floor removes whole payments. Size 3, January 2026, gross wages $2,054 a month: head sc_tanf_gross_income_eligible = False and sc_tanf = 0. Main is True and pays 33.75. The head's limit is floor(1.85 × 1,110) = 2,053; main's is 1.85 × 1,110.42 = 2,054.27.

So expect a few households to lose their whole benefit at the gross limit, and most to lose up to about $1.40. The body makes no claim about the size of the loss, so the body needs no edit.

P3. Mutant 2 is killed only in January; Cases 21 and 22 can't distinguish the order

PR body, mutant table, row 2.

A $100 flat disregard commutes with subtracting the support whenever earnings exceed S + 100, so the May cases give the same countable income in either order. I wrote mutant 2 as "subtract S from gross earnings, then apply the §8.12 disregard". integration.yaml killed it only in Case 23 (May countable 100, not 0, because the $100 floor bites) and Case 25 (January countable 200, not 100). It survived Cases 21 and 22 and every property. The hand-off's mutant was weaker, as it says itself, so its list of killing cases describes that mutant, not the ordering rule. Replace the row's "Killed by" cell with: "integration.yaml Cases 23 and 25. Case 25 (January) is the ordering case: in May the $100 disregard commutes with the deduction."

P3. Body table: main's January value for Case 25 is 327.89, not 327.90

PR body, Expected values, row "Parent earns $600/mo and pays $2,400/yr, 2026-01". Main gives 327.89456 ((1,110.4167 − 300) × 0.4046 = 327.8946). The hand-off comment has it right. Replace 327.90 with 327.89.

P3. Invariant wording: (b) and (f) say more than the tests check

• (b) says "Countable income is at least 0." With a business loss, countable income is negative on main and at this head alike. Head, January 2026, parent with −$14,400 a year of self-employment income: sc_tanf_countable_income = −600 and sc_tanf = 356 (capped at the payment standard). The body qualifies (b) two paragraphs later. State the condition in (b) itself: "(b) With no self-employment loss, countable income is at least 0."
• (f) says countable income "equals main's formula exactly". The test uses assert_allclose(..., atol=0.01) against a frozen function. I read the frozen function against main's sc_tanf_countable_earned_income_person and sc_tanf_countable_income and it matches for the generated input space (no SSI, no self-employment, no interest). Replace "exactly" with "to the cent".

P3. sc_tanf.py still cites Volume 65

policyengine_us/variables/gov/states/sc/tanf/sc_tanf.py:13 keeps tanf-policy-manual-volume-65.pdf#page=131 under the "PDF pages 131, 133" comment, and the PR adds Volume 67 p.133 beside it. The repository's reference rule is one entry per source document. Since every other file in the diff moved to Volume 67, drop the Volume 65 line, or confirm that the page-131 citation still means Volume 65 and fix the comment.

Checked and sound

• The defect at main. Main counts a 16-year-old's $500 a month: Case 19's household gives countable 250 and sc_tanf 348.12 at origin/main. At the head it gives countable 0 and sc_tanf 449.
• Hand values, computed before reading the code. 2025 guideline: $15,650 + $5,500 per added person.
  • Need ⌊FPG/24⌋ for sizes 1–8: 652, 881, 1,110, 1,339, 1,568, 1,797, 2,027, 2,256.
  • Gross limit ⌊1.85 × need⌋: 1,206, 1,629, 2,053, 2,477, 2,900, 3,324, 3,749, 4,173.
  • Payment ⌊0.4046 × need⌋: 263, 356, 449, 541, 634, 727, 820, 912.
  • Size 9: 59,650/24 = 2,485.42, so 2,485, which is 2,256 + 229.
  • The household cases: 449; 449 (main 348.12); 327 (main 246.97); 327 (payer-only reading 246); 449 (main 348.12); 408 (main 327.89); 12 (main 12.81); 0 (main 10.06).
  • Every value matches the body, the YAML and a run. The size-8 table cells are pinned at 2026-01 in sc_tanf_need_standard.yaml, sc_tanf_gross_income_eligible.yaml and sc_tanf.yaml.
• Integration Cases 21, 22, 23 and 25 by hand.
  • Case 21, May: E = 600 − 100 = 500, S = 2,400/12 = 200, countable 300, gross 600, floor(810 × 0.4046 = 327.73) = 327.
  • Case 22: the same group income comes from adult B, and A's support of 200 comes off it, so 327. Limiting the deduction to the payer's own earnings leaves countable 500 and floor(610 × 0.4046 = 246.81) = 246.
  • Case 23: E = 150, U = 100, S = 300, deduction min(300, 250) = 250, countable 0, gross 350, payment 449.
  • Case 25, January: E = 300, deduction 200, countable 100, floor(1,010 × 0.4046 = 408.65) = 408.
• Existing expectations. The rewritten sc_tanf.yaml cases check by hand: 344.31 → 344; 142.01 → 142; 4.45 → 4, unpaid; 12.54 → 12; 9.72 → 9, unpaid; 10.03 → 10, paid; 9.79 → 9, unpaid. The payment rate is 0.4046 from 2024-10-01.
• Rounding code. The need standard, the gross limit and the payment are each floored, and the payment is floored before the $10 test. need_standard/rate.yaml has a single value, 0.5 from 1996-07-01, so the ⌊⌊x⌋/2⌋ = ⌊x/2⌋ comment holds at every date the code reads (x ≥ 0). Countable income isn't rounded.
• Mutant 6 is equivalent. minimum_payment is the integer 10, and for an integer n, ⌊x⌋ ≥ n exactly when x ≥ n. I ran it anyway: results below.
• Above eight persons.
  • The head gives need 2,485 (size 9), 3,172 (12), 3,402 (13), 4,547 (18) and 4,777 (19) for 2026-01. Main gives 3,402.08 for 13.
  • The manual's rule (2,256 + 229k) gives 3,172, 3,401, 4,546 and 4,775. So the two agree through 12, the regulation is $1 higher for 13–18 and $2 higher for 19–24 (size 25 is +3), as the candidate entry says.
  • The entry's example (81,650/24 = 3,402.08) and its note (5,500/24 = 229.17) check.
  • test_source_conflicts.py passes (103). The fields follow the house format for an unregistered candidate. Like other candidates, it carries no # Source conflict marker, which the test forbids until the entry is registered.
  • No other SC TANF parameter or test assumes $229. The only mention is the size-9 comment in sc_tanf_need_standard.yaml.
• The October revision. For a size-3 group, the head gives 1,110 in 2026-05 and 2026-09 (the 2025 guideline) and 1,138 in 2026-10 (the 2026 guideline).
• The May-input claim. A copy of Case 21 written with period: 2026-05 fails, with deduction 0 against 200: the runner stores the yearly inputs where May can't see them. The PR's period: 2026 with month-keyed outputs avoids this. It hides nothing in the model, but no Oct–Dec month is tested for the support cases.
• Dependent child, apart from P1 and P2:
  • An 18-year-old with a diploma (not in school) isn't a dependent child, and their $500 counts (sc_tanf 347).
  • An 18-year-old full-time in primary school is treated as a non-student. The body discloses this.
  • A minor parent who is a tax dependent of her own parent is excluded (sc_tanf_dependent_child.yaml Case 4).
  • A minor parent with an explicit head role counts (Case 2).
  • A minor parent heading her own tax unit inside a parent's SPM unit counts: the formula reads only her tax-unit role.
  • Pregnancy plays no part: the formula never reads is_pregnant, and Case 3 and integration Case 24 pin it.
  • A dependent child's unearned income still counts (integration Case 20: 200 counted, 368).
• Child support.
  • Payers are members who aren't SSI recipients and aren't dependent children.
  • The deduction is min(S, max(E + U, 0)), with E the sum of sc_tanf_countable_earned_income_person, which is after the disregards.
  • sc_tanf_gross_income still only adds gross earned and gross unearned income, and mutant 3 is killed.
  • A group with no income gets no deduction.
  • child_support_expense is definition_period = YEAR, and the monthly read divides it by 12 (Case 21: 2,400 → 200).
• Negative income.
  • Parent with −$14,400 a year of self-employment and $2,400 a year of support, January 2026: deduction 0, countable −600 on both, sc_tanf 356. In May the $100 disregard floors each person at 0.
  • −$1,200 a year of self-employment with $3,600 of alimony, January: E = −50, U = 300, deduction 200, countable 50. That equals the stated allocation max(E − S, 0) + max(U − max(S − E, 0), 0) = 0 + (300 − 250).
  • The zero floor in sc_tanf_countable_income applies only when the deduction is positive. A positive deduction implies positive income, and the result is at most income − deduction plus float error. So the floor can't raise a payment above main's except by that rounding error.
• Readers.
  • sc_tanf_gross_earned_income is read by sc_tanf_gross_income and sc_tanf_countable_earned_income_person.
  • sc_tanf_countable_income is read by sc_tanf_countable_income_eligible (and through it sc_tanf_income_eligible) and sc_tanf.
  • sc_tanf_need_standard is read by sc_tanf_gross_income_eligible, sc_tanf_countable_income_eligible and sc_tanf.
  • sc_tanf is read by tanf_if_takes_up, then tanf. From tanf it reaches tanf_person, spm_unit_benefits (and through it SPM net income and poverty), hud_countable_unearned_income, meets_wic_categorical_eligibility and is_head_start_categorically_eligible. The other states' readers of tanf (NH, RI, OR) don't see South Carolina households.
  • All of these move at baseline for South Carolina households.
• Overlap. I listed all 179 open PRs and their files. None touches states/sc/tanf. Fifteen edit source_conflicts/candidates.yaml. In a trial merge with git merge-tree, thirteen merge cleanly and Add South Carolina LIHEAP regular heating assistance #10616 (SC LIHEAP) conflicts in that file only, because both add us-sc: entries at the same place. For Include farm and partnership earnings in SSI and correct loss netting #10532 and Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 the trial merge returned an error, so they're unchecked. Nothing in Michigan or South Carolina TANF and Michigan FIP: whole-dollar rounding next to the $10 minimum is not applied (114-1140(A)(4), BEM 520) #10497's other half is touched.
• Scope. Nothing under .github/ or policyengine_us/tests/policy/baseline/partners/ changes, and no partner test reads SC TANF. The Household API Partners job passed. The changelog fragment is a top-level changelog.d/*.fixed.md. The body has the impact: and axiom: lines.
• Tests at this head, one file per process:
  • The 15 SC TANF YAML files, 128 cases: integration 25, sc_tanf 16, gross_income_eligible 21, gross_unearned 18, need_standard 12, dependent_child 4, countable_earned_person 4, countable_income_eligible 4, eligible 4, income_eligible 4, resources_eligible 4, ssi_recipient 4, benefit_group_size 3, countable_resources 3, gross_earned 2.
  • test_sc_tanf_child_earnings_support_invariants.py: 4 passed, 45 s wall with startup.
  • test_sc_tanf_unearned_income_invariants.py: 2 passed.
  • code_health/test_source_conflicts.py: 103 passed.
  • The counts match the body.
• Properties. None passes trivially: mutants 1 and 3 each fail test_invariants. Gaps:
  • (c) varies only dependent-child wages. No property checks that a head-of-household minor's wages do count, which is the P1 gap.
  • (d) raises only member 0's wages.
  • The ranges reach 0 and support above income, but max_examples=10.
  • No property targets the order of the deduction and the disregard.
• Mutants. Each ran in its own process in a scratch worktree. The worktree's git status was clean afterwards. For each mutant I ran integration.yaml, sc_tanf_dependent_child.yaml, sc_tanf_need_standard.yaml, sc_tanf.yaml, sc_tanf_gross_income_eligible.yaml, sc_tanf_countable_income_eligible.yaml and the property file.

#	Mutant	Result	Killed by
1	Drop the dependent-child exclusion	killed	integration Cases 19, 20; dependent_child Case 4; test_invariants (c)
2	Subtract S from gross earnings, then apply the disregard	killed	integration Cases 23, 25 only (see P3)
3	Child support reduces sc_tanf_gross_income	killed	integration Cases 21, 23, 25; test_invariants (e)
4	Deduction capped at each payer's own countable earnings	killed	integration Cases 22, 23
5	No need-standard floor	killed	need_standard 11 of 12; integration 6; gross_income_eligible 8; sc_tanf 1; countable_income_eligible 1
6	$10 test on the unfloored benefit, floor after	survived, equivalent	none, as expected: the threshold is an integer
7	(age test & ~head_or_spouse) | is_pregnant	killed	dependent_child Case 3; integration Case 24; both (f) differentials
8 (mine)	Drop ~head_or_spouse	killed	dependent_child Case 2 only

Mutant 8 is caught by a single case, and that case inputs the head role explicitly. That gap is P1.

The P1 fix above, applied in a scratch tree, passes sc_tanf_dependent_child.yaml (4), integration.yaml (25) and the property file (4).

What I couldn't check

• Sources. dss.sc.gov and scstatehouse.gov are outside this session's network allowlist. Both a proxy fetch and a direct fetch failed, so I didn't try other hosts. I couldn't read §4.10, §7.5, §8.11, §8.15, §8.16 or 114-1140 myself. Every quotation above is the body's or the hub's, and none is confirmed. The rule logic is checked against those quotations only.

• rulespec-us#1728. That repository is outside this session's GitHub scope, so I couldn't check its seven companion rows against this head's YAML, or whether the issue is labelled pe-parity and dispatch-ready as queued requires. The body's line reads axiom: TheAxiomFoundation/rulespec-us#1728 queued.

• Microdata. I couldn't check whether the default dataset supplies tax_unit_role_input for minors, which decides how far P1 and P2 reach in microsimulation. I also couldn't check how often child_support_expense is nonzero for SC TANF units.

• Population. I ran no microsimulation. For the impact pair at the landing head, beyond sc_tanf and the common set, the hub needs:

  • sc_tanf_dependent_child
  • sc_tanf_gross_earned_income
  • sc_tanf_gross_income
  • sc_tanf_child_support_deduction
  • sc_tanf_countable_income
  • sc_tanf_need_standard
  • sc_tanf_gross_income_eligible
  • sc_tanf_countable_income_eligible
  • sc_tanf_eligible
  • child_support_expense (input coverage)
  • tanf, spm_unit_benefits and the SPM poverty flags

  Expected directions: the child-earnings exclusion and the support deduction raise payments and add eligible units. The floors lower most payments by up to about $1.40, and the gross-limit floor drops a few units at the limit (P3).

• CI. I looked at CI only at this head.

---

Generated by Claude Code

[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]

Picking this up for the US hub. I continue an earlier session's round 2 work on this PR (its four commits, 0beeace6d6 to b58a4031a1), which stopped without a hand-off. I'll post a hand-off comment here when it's ready.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
71e1d42

Round 2 fix round, following the round 1 review (#10581 (comment)). This session continued an earlier session's work. That session pushed four commits (0beeace6d6 to b58a4031a1) and posted no hand-off. I checked each against the brief before building on it, and kept all four. I added one more merge of main. This round's own changes are that merge and the PR body.

The earlier session's commits

• 0beeace6d6, a merge of main: kept. It merged 9a494128, which is on main. Superseded by 71e1d42269 below.
• 0da9f14e8b, finding 1: kept. sc_tanf_dependent_child.py:33-39 is the review's code, character for character (no_adult, oldest = person.get_rank(person.spm_unit, -age) == 0, heads_household). documentation at :14-20 is the brief's text, word for word. The new cases are sc_tanf_dependent_child.yaml Case 5 (:89-108) and integration.yaml Case 26 (:780-805), period 2026-01, with no role inputs.
• b65a975fe0, finding 2: kept. integration.yaml Case 27 (:807-836): spouse 20, wife 17 with is_tax_unit_spouse: true and $6,000 of wages, and a baby. It declares a couple marital unit and a separate one for the baby, because with no marital units declared everyone shares one, which the review flagged.
• b58a4031a1, finding 7: kept. The sc_tanf.py:11-15 reference tuple is Vol. 67 p.133 and 114-1140 p.210, and Vol. 65 is gone. One judgment call: the comment reads # PDF pages 133-134, not the brief's # PDF pages 133, 134. CLAUDE.md asks for a range for consecutive pages ("#page 72-75 (consecutive)"), and sc_tanf_need_standard.py:11 already uses 133-134. If the hub wants the brief's form, it's a one-character change.

Merge of main

71e1d42269 merges origin/main at 89a15b72f5 with no conflicts. source_conflicts/candidates.yaml changed on both sides and auto-merged. test_source_conflicts.py passes 103 after the merge. git diff origin/main HEAD touches only the PR's 17 files.

A setup note: the clone came shallow. My first merge attempt saw 97b6d513 as the merge base and reported conflicts in CHANGELOG.md, pyproject.toml, uv.lock and mo_itemized_deductions.yaml. Those were artifacts of the shallow history. I aborted that merge, deepened the history (merge base 9a494128, as GitHub's compare reports), and the real merge was clean.

Hand values (new cases)

• Case 26 and dependent_child Case 5, mother 17, $6,000 a year, baby 0, 2026-01, no roles:
  • sc_tanf_dependent_child [false, true], gross earned [500, 0].
  • Need for 2: floor((15,650 + 5,500) / 12 × 0.5 = 881.25) = 881.
  • Countable: 500 × 0.5 = 250.
  • floor((881 − 250) × 0.4046 = 255.30) = 255. At cdf464ad90 it was 356 (the review's witness), and main gives 255.40.
• Case 27, 2026-01:
  • sc_tanf_dependent_child [false, false, true], countable 250.
  • floor((1,110 − 250) × 0.4046 = 347.96) = 347. Main gives 348.12, from the review's run.
• Both pass at this head.

Findings 1–8

1. Code, documentation and both cases: see above.
2. The case is in place (above). The body's Mapping paragraph carries the brief's replacement sentence word for word, and the old "almost always" sentence is gone.
3. There's no "Status: draft" line. "Tests" now ends with the CI run at this head (below).
4. The mutant 2 row reads "integration.yaml Cases 23 and 25. Case 25 (January) is the ordering case: …". It's backed by this round's run (a below).
5. The Expected values row "Parent earns $600/mo and pays $2,400/yr, 2026-01" reads 327.89.
6. (b) reads "(b) With no self-employment loss, countable income never falls below 0." (f) reads "to the cent", and the paragraph after the table reads "so (f) still holds".
7. See b58a4031a1 above.
8. The body was written from a file with PATCH, and GitHub's readback matched the file, apart from the trailing newline jq adds. It has mutant rows 8–10, this round's counts, the Overlap section, the brief's impact: line, and the unchanged axiom: line and closing line.

Mutants

Each ran in its own process at 71e1d42269, whose SC TANF code is identical to b58a4031a1. Each ran against integration.yaml, sc_tanf_dependent_child.yaml, sc_tanf.yaml, sc_tanf_need_standard.yaml, sc_tanf_gross_income_eligible.yaml, sc_tanf_countable_income_eligible.yaml, sc_tanf_gross_earned_income.yaml and the property file. The source was restored from a backup after each, and git status was clean.

Mutant	Result	Failing tests
(a) mutant 2: each earner's share of S off gross earnings (floored at 0), then the per-person §8.12 disregard	killed	integration.yaml Case 23 and Case 25 only; all four property tests pass
(a′) cruder group version: S off group gross earnings, one $100 disregard per group, no floor	killed	integration.yaml Cases 23, 25; property test_invariants (b: January countable −0.04 when support > earnings), test_matches_previous_formula… and the May seeded-households test. Those property failures come from the group-level disregard and the missing floor, not from the ordering, so the body's row uses (a)
(b) heads_household = head_or_spouse	killed	sc_tanf_dependent_child.yaml Case 5; integration.yaml Case 26
(c) oldest → True	killed	sc_tanf_dependent_child.yaml Case 5 (the baby) and Case 2; integration.yaml Case 26
(d) return meets_age_test	killed	sc_tanf_dependent_child.yaml Cases 2, 5; integration.yaml Cases 26, 27

No property test kills (b), (c) or (d), because every generated group has an adult head.

Tests at 71e1d42269 (one file per process)

The SC TANF YAML files, 131 cases, all passing:

• integration 27
• gross_income_eligible 21
• gross_unearned_income 18
• sc_tanf 16
• need_standard 12
• dependent_child 5
• countable_earned_income_person 4
• countable_income_eligible 4
• eligible 4
• income_eligible 4
• resources_eligible 4
• ssi_recipient 4
• benefit_group_size 3
• countable_resources 3
• gross_earned_income 2

Python, all passing:

• test_sc_tanf_child_earnings_support_invariants.py 4 (34 s wall, local)
• test_sc_tanf_unearned_income_invariants.py 2
• code_health/test_source_conflicts.py 103
• test_known_dead_reference_urls.py 40
• test_reference_urls.py 32
• test_no_concatenated_reference_urls.py 48

ruff format --check and ruff check are clean on the nine changed Python files (ruff 0.15.5). No microsimulation was run, and no partner or .github/ file was touched.

Overlap (git merge-tree against each PR's current head)

• Add South Carolina LIHEAP regular heating assistance #10616 (d2246f17d7): it merges cleanly into main, but conflicts with this PR in source_conflicts/candidates.yaml only. us-sc:liheap-income-floor-vs-smi-limit and us-sc:tanf-need-standard-above-eight land in the same place. Whichever PR lands second keeps the file sorted.
• Include farm and partnership earnings in SSI and correct loss netting #10532 (4cab6b38ae): merges cleanly with this PR.
• Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 (5a92f6cc0c, a draft on state-eitc-section-911-md-me-nj): it already conflicts with main in Makefile, me_eitc.py and one candidates.yaml hunk near line 4250. Against this PR it has the same conflicts, and this PR's entry (line 8441 of the trial merge) merges cleanly. So this PR adds no conflict there.

CI

Pull request run 41714176878 at 71e1d42269 was still in progress at 21:42Z, 65 minutes after the 20:37Z push, so the run has no conclusion yet. 36 of 37 jobs had completed with success: every Baseline shard, every Rest Python shard, Microsimulation, every other Contrib shard, Household API Partners, PR validation, Package and compatibility, and Quick Feedback. Contrib other-shard-3 was still running. The previous head b58a4031a1 passed run 41538838024.

Judgment calls

• # PDF pages 133-134 rather than the brief's 133, 134 (see above).
• I kept Case 27's explicit marital units.
• Main's values for Cases 26 and 27 in the body table come from the round 1 review's runs on main. I didn't rerun main this round.
• The brief cites sc_tanf_dependent_child.py "lines 33-34". After the earlier session's commit the fix spans lines 33-39, because the added comment sits inside it.

Proposed follow-up

None.

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

Round 2 review at 71e1d4226917c7690a8e046b6d9decb53b090877 (the head hasn't moved).

Firm so far:

• CI: Pull request run 41714176878 at this head concluded success, 36 of 36 jobs. The body's CI sentence is stale (P3).
• All 15 SC TANF YAML files pass (131 cases). The property file passes 4, the unearned invariants file 2, test_source_conflicts.py 103 and test_known_dead_reference_urls.py 40.
• Hand values for the new cases match: 255 (main 255.40) and 347 (main 348.12). The defect reproduces at main: a 16-year-old's $500 counts and pays 348.12, against 449 at the head.
• Likely P2, the tie. In a group with no adult whose oldest members tie in age, get_rank (a stable sort) picks whoever is listed first, so the result depends on input order. Twins aged 17 and a baby, one twin earning $500 a month, 2026-01: the earner listed first gives 347, listed second gives 449. Main gives 348.12 either way.
• P3: mutant heads_household = head_or_spouse | oldest (drops no_adult &) survives every SC test.
• Overlap: Add South Carolina LIHEAP regular heating assistance #10616's head moved to 85d8103fbd and still conflicts only in candidates.yaml; it now adds six us-sc:liheap-* entries, all of which sort before us-sc:tanf-need-standard-above-eight.

Mutant case names and the fix check are still running.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: REQUEST_CHANGES
71e1d42
METHODOLOGY_CHOICE: yes

Independent round 2 review. The head hasn't moved past 71e1d42269. Current main (adea7035ea) is five commits past the merged 89a15b72f5. None of those commits touches SC TANF, tax-unit roles or is_child, and git merge-tree merges this head into it cleanly.

METHODOLOGY_CHOICE stays yes. The PR keeps round 1's readings: tax-unit head or spouse stands in for "head-of-household", one member's support comes off the whole group's income, and the regulation governs above eight persons. It adds the hub's reading that the oldest member heads a group with no adult.

Every round 1 finding is fixed at this head. One new defect is left in the no-adult clause: when the oldest members tie in age, the result depends on the order the members are listed.

Findings

P2. When the oldest members of a group with no adult tie in age, the result depends on input order

policyengine_us/variables/gov/states/sc/tanf/sc_tanf_dependent_child.py:40

oldest = person.get_rank(person.spm_unit, -age) == 0 gives rank 0 to exactly one member. Core's get_rank sorts with kind="stable", so in a tie the member listed first wins. The other tied member is a dependent child, and her earnings drop out. The hub's ruling says the oldest member heads the household, but it doesn't say which of two equally old members that is. The output shouldn't depend on which one the input lists first. Main counts her wages in either order, so this PR makes the case reachable and makes it worse.

Witness, 2026-01, no role inputs: twins aged 17, one earning $6,000 a year, and a baby aged 0.

uv run python -I probe.py     # Simulation(situation=...), explicit SPM, tax, family and household units, state SC
head, earner listed first:  sc_tanf_dependent_child=[F, T, T]  gross earned=[500, 0, 0]  countable=250  sc_tanf=347
head, earner listed second: sc_tanf_dependent_child=[F, T, T]  gross earned=[0, 0, 0]    countable=0    sc_tanf=449
main, either order:         countable=250  sc_tanf=348.12

By hand: size 3 need ⌊(15,650 + 2 × 5,500) / 12 × 0.5 = 1,110.42⌋ = 1,110. Her $500 less the 50% January disregard leaves 250, so the payment is ⌊(1,110 − 250) × 0.4046 = 347.96⌋ = 347. With her wages dropped it's ⌊1,110 × 0.4046 = 449.11⌋ = 449.

Fix. Every member who ties for oldest heads the household. Two members of the same age can't be each other's parent, so neither should be the other's dependent child. Replace lines 36-41 with:

        # Tax-unit roles are inferred only for adults, so a minor parent who
        # lives with no adult heads no tax unit unless roles are input. The
        # oldest member of a group with no adult heads the household; members
        # who tie for oldest all head it, so the result doesn't depend on the
        # order in which members are listed.
        no_adult = ~person.spm_unit.any(~person("is_child", year))
        oldest = age == person.spm_unit.max(age)
        heads_household = head_or_spouse | (no_adult & oldest)

I applied that change in place and restored it afterwards (git status clean). It gives 347 in both orders, with sc_tanf_dependent_child [F, F, T] and the earner's $500 counted. The no-adult witness still gives 255, and a 16-year-old mother with a 17-year-old sibling still gives 449, as the ruling's "oldest member" requires. integration.yaml (27) and sc_tanf_dependent_child.yaml (5) still pass, and mutants (b) to (e) below still have killers, since Case 5's baby is younger than the mother.

Add a case to sc_tanf_dependent_child.yaml, and the same household to integration.yaml, at period: 2026-01 with no role inputs. People in this order: sibling aged 17; minor_parent aged 17 with employment_income_before_lsr: 6_000; baby aged 0. Expect sc_tanf_dependent_child: [false, false, true] and sc_tanf_gross_earned_income: [0, 500, 0], and in integration.yaml sc_tanf_countable_income: 250 and sc_tanf: 347 (⌊(1,110 − 250) × 0.4046 = 347.96⌋). The case fails at this head (449) and passes with the fix.

Then update the variable's documentation, replacing "in a group with no adult the oldest member does" with "in a group with no adult the oldest member does, and members who tie for oldest all do". In the body's Mapping paragraph, replace "In a group with no adult, the oldest member heads the household." with "In a group with no adult, the oldest member heads the household, and members who tie for oldest all head it."

Reach: household calculations only, as with round 1's P1. The hub reports that the default dataset supplies every person's tax-unit role and that every member of its no-adult SPM units is a tax-unit head, so head_or_spouse already covers them there. I couldn't check that data myself.

P3. The body's CI sentence is stale

PR body, "Tests", last paragraph. It says run 41714176878 "was still in progress 65 minutes after the push", with Contrib other-shard-3 still running. The run has since finished: gh api repos/PolicyEngine/policyengine-us/actions/runs/41714176878/jobs lists 36 jobs, all success, with other-shard-3 completing at 21:43:11Z. The hand-off's "36 of 37" overcounts by one. Replace the sentence with: "CI: Pull request run 41714176878 at 71e1d42269 concluded success; all 36 jobs passed." If the P2 fix lands, report the run at that head instead.

P3. The body's #10616 overlap bullet describes an older head

PR body, "Overlap". #10616's head is now 85d8103fbd, not the d2246f17d7 the hand-off used. It still conflicts with this PR only in source_conflicts/candidates.yaml, and still merges cleanly into main. But it now adds six us-sc: entries, not one: liheap-energy-burden-test, liheap-income-floor-vs-smi-limit, liheap-low-income-tier-base, liheap-minor-income, liheap-mixed-status-household and liheap-plan-income-checklist. The conflict hunk opens at us-sc:tanf-need-standard-above-eight against us-sc:liheap-energy-burden-test. The file is sorted by id: 358 entries at this head, ids == sorted(ids). So all six LIHEAP entries go before this PR's TANF entry. Replace that bullet's last two sentences with: "Both add us-sc: entries after the last us-ri: entry: six us-sc:liheap-* entries there, us-sc:tanf-need-standard-above-eight here. Whichever PR lands second keeps the file sorted by id, with all six LIHEAP entries before the TANF entry."

Checked and sound

• Round 1 findings at this head:

  • P1: the review's code is in place at sc_tanf_dependent_child.py:35-42, and the documentation is the brief's text.
  • P2: the Mapping paragraph carries the brief's replacement, and Case 27 pins the role-input path.
  • P2 status: the "Status: draft" line is gone. Only the CI sentence is stale (P3 above).
  • Mutant 2 row: matches my run, below.
  • Main's January Case 25 value: reads 327.89. (1,110.4167 − 300) × 0.4046 = 327.8946.
  • (b) and (f) wording: matches the brief.
  • sc_tanf.py reference: Volume 67 p.133 plus Chapter 114 p.210; Volume 65 is gone. # PDF pages 133-134 follows CLAUDE.md's range form.

• The ruling. With no role inputs, a 17-year-old mother alone with her baby counts as a parent: sc_tanf_dependent_child [F, T], gross earned [500, 0], sc_tanf 255. With an adult in the group, the head-or-spouse rule stands: a 17-year-old married to a 20-year-old, with no role inputs, is a dependent child and pays 449, as the body discloses. Input as spouse, she counts and pays 347. Ties are the P2 above.

• Hand values, computed before reading the YAML. Parameters at 2026-01:

  • guideline at instant 2025-10-01: $15,650 + $5,500;
  • need rate 0.5; payment rate 0.4046 from 2024-10-01;
  • 50% disregard in months 1-4, $100 after.

  Results:

  • Case 5 and Case 26: need ⌊881.25⌋ = 881, countable 250, ⌊631 × 0.4046 = 255.30⌋ = 255. Main gives 255.40.
  • Case 27: need 1,110, countable 250, ⌊860 × 0.4046 = 347.96⌋ = 347. Main gives 348.12.
  • Both main values come from my run at origin/main. Both match the body table and the YAML.
  • Further probes at the head, by hand and by run:
    • no-adult minor parent paying $1,200 a year of support: deduction 100, countable 150, ⌊731 × 0.4046 = 295.76⌋ = 295;
    • the same minor parent in May: 500 − 100 = 400, ⌊481 × 0.4046 = 194.61⌋ = 194.

• The original defect. Reproduced at origin/main: parent 35, child 16 earning $6,000 a year, child 8, 2026-01. Main gives countable 250 and sc_tanf 348.12. The head gives countable 0 and 449.

• Tests at this head, one file per process:

  • YAML: the 15 SC TANF files, 131 cases, all passing. integration 27, gross_income_eligible 21, gross_unearned_income 18, sc_tanf 16, need_standard 12, dependent_child 5, countable_earned_income_person 4, countable_income_eligible 4, eligible 4, income_eligible 4, resources_eligible 4, ssi_recipient 4, benefit_group_size 3, countable_resources 3, gross_earned_income 2.
  • Python: test_sc_tanf_child_earnings_support_invariants.py 4 (52 s wall, needs the dev extra for Hypothesis), test_sc_tanf_unearned_income_invariants.py 2, test_source_conflicts.py 103, test_known_dead_reference_urls.py 40, test_reference_urls.py 32, test_no_concatenated_reference_urls.py 48.
  • These match the body.

• New tests assert the claim.

  • Case 5 asserts the role (sc_tanf_dependent_child) and its effect (sc_tanf_gross_earned_income), with no role inputs.
  • Case 26 asserts gross, countable, need and benefit.
  • Case 27 asserts the spouse's role and the payment.
  • None covers a tie (P2).

• Mutants. Each ran in place in its own process and was restored with git checkout, and git status was clean after each. Each ran against integration.yaml, sc_tanf_dependent_child.yaml, sc_tanf.yaml, sc_tanf_gross_earned_income.yaml, sc_tanf_countable_income_eligible.yaml and the property file.

Mutant	Result	Killed by
(a) each earner's share of S, by gross earnings, off gross earnings floored at 0, then the per-person §8.12 disregard	killed	integration.yaml Cases 23, 25; all four properties pass
(b) heads_household = head_or_spouse	killed	sc_tanf_dependent_child.yaml Case 5; integration.yaml Case 26
(c) no_adult & True (oldest → True)	killed	sc_tanf_dependent_child.yaml Cases 2, 5 (the baby); integration.yaml Case 26
(d) return meets_age_test	killed	sc_tanf_dependent_child.yaml Cases 2, 5; integration.yaml Cases 26, 27
(e, mine) the youngest member heads: get_rank(person.spm_unit, age) == 0	killed	sc_tanf_dependent_child.yaml Cases 2, 5; integration.yaml Case 26
(f, mine) drop no_adult &: head_or_spouse | oldest	survived, equivalent for household input	In a group with an adult the oldest member is an adult, and meets the age test only at 18 in secondary school. Inference then makes them head, even with is_tax_unit_dependent: true (checked: is_tax_unit_head [T, F] for an 18-year-old student input as a dependent, with a 10-year-old). It could differ only with dataset-supplied roles.

My first version of (a) failed all four properties too. That was my bug: an unprojected SPM-level sum broke multi-unit simulations. The corrected version gives the result above, which matches the body's row 2. Rows 8-10 of the body's table match (b)-(d) exactly.

No property test kills (b)-(e), because every generated group has an adult head. I don't think a generator for groups with no adult is worth adding. YAML Cases 2, 5, 26 and 27 already kill every mutant of the clause, and the tie case above pins the remaining gap more cheaply than a Hypothesis strategy that builds simulations.

• Readers.
  • sc_tanf_dependent_child → sc_tanf_gross_earned_income and sc_tanf_child_support_deduction (payer test).
  • sc_tanf_child_support_deduction → sc_tanf_countable_income.
  • sc_tanf_gross_earned_income → sc_tanf_gross_income, sc_tanf_countable_earned_income_person.
  • sc_tanf_countable_income → sc_tanf_countable_income_eligible (→ sc_tanf_income_eligible), sc_tanf.
  • sc_tanf_need_standard → sc_tanf_gross_income_eligible, sc_tanf_countable_income_eligible, sc_tanf.
  • sc_tanf_gross_income_eligible → sc_tanf_income_eligible.
  • sc_tanf → tanf_if_takes_up → tanf, and on from there as round 1 listed.
  • In a no-adult group the oldest member now also counts as a support payer: the $1,200-a-year probe gives a deduction of 100. That's consistent with the ruling.
  • Beyond the P2, no reader is made worse.
• Baseline movement. The diff moves SC TANF outputs at baseline: the child-earnings exclusion, the support deduction, and the three floors. The round 2 change, the no-adult clause, should move only household calculations if the hub's dataset reading holds. The body's impact: line exists and states the expected directions. The hub's population pair should cover sc_tanf, sc_tanf_dependent_child, sc_tanf_child_support_deduction, sc_tanf_countable_income, sc_tanf_gross_income_eligible and tanf.
• Body lines. impact: and axiom: TheAxiomFoundation/rulespec-us#1728 queued both exist.
• Overlap.
  • A search for open PRs on "sc tanf" returned 12 besides this one. Only Round summed resources to whole cents before exact-limit asset tests #10639 touches an SC TANF file (sc_tanf_resources_eligible.yaml, which this PR doesn't change), and it merges cleanly with this PR.
  • Include farm and partnership earnings in SSI and correct loss netting #10532 (4cab6b38ae) merges cleanly with this PR. It now conflicts with main, which this PR doesn't cause.
  • Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 (5a92f6cc0c, draft) conflicts in Makefile, me_eitc.py and one candidates.yaml hunk at us-me:childless-eitc-age-cap, the same against main as against this PR. This PR's entry merges into it cleanly.
  • Add California LIHEAP regular heating assistance #10607, Add Colorado LEAP regular heating assistance #10609, Add Connecticut CEAP regular heating assistance #10611, Add Delaware LIHEAP regular heating assistance #10612, Add Idaho LIHEAP regular heating assistance #10613, Add Hawaii LIHEAP regular heating eligibility and subsidized-dwelling benefit #10614, Fix the 2020 unemployment compensation exclusion: cutoff test, per-spouse cap, dependents #10175 and Apply 416.1163(e)(2) to deemed SSI and identify SSI couples by marriage, not tax roles #10141 merge cleanly with this PR.
• Scope. Nothing under .github/ or policyengine_us/tests/policy/baseline/partners/ changes. The changelog fragment is a top-level changelog.d/sc-tanf-child-earnings-support-whole-dollars.fixed.md.
• CI. Run 41714176878 at this head concluded success, 36 of 36 jobs.

What I couldn't check

• Sources. From this session, dss.sc.gov fails DNS resolution, and scstatehouse.gov, govinfo.gov, ecfr.gov, uscode.house.gov and web.archive.org return a proxy 403. I couldn't read Policy Manual Vol. 67 §§4.10, 7.5, 8.11, 8.15 or 8.16, or S.C. Code Regs. 114-1140. Every quotation in the body is unconfirmed by me, and I checked the rules only against the quotations in the body and the hub's brief. The no-adult reading rests on the hub's ruling, not on a source I read.
• Microdata. I couldn't read the default dataset, so the claim that every member of a no-adult SPM unit has a supplied HEAD role, which limits this round's change to household calculations, is the hub's, unverified here.
• Population impact. I ran no microsimulation.
• rulespec-us#1728. That repository is outside this session's GitHub scope, so I couldn't check that the issue is labelled pe-parity and ready to dispatch.

---

Generated by Claude Code

[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
c671a31

Round 2 review: one P2 and two P3s, all addressed. Nobody else committed to the branch while I worked. It was at 71e1d42269 when I started and before I pushed.

1. Pick-up and merge

• Pick-up comment posted.
• The branch was 8 commits behind origin/main. I merged it with a merge commit (c671a313e3), with no conflicts; source_conflicts/candidates.yaml auto-merged. I reran all the changed files after the merge (below).

2. P2: members who tie for oldest all head the household

• Code: sc_tanf_dependent_child.py:36-43 carries the review's code verbatim. The surrounding names (age, year, head_or_spouse) matched, so I didn't adapt anything.
• Documentation: sc_tanf_dependent_child.py:17-19 now reads "...in a group with no adult the oldest member does, and members who tie for oldest all do, so a minor parent...".
• Cases, all at 2026-01 with no role inputs:
  • sc_tanf_dependent_child.yaml:110 Case 6 lists sibling, minor_parent, baby. It expects sc_tanf_dependent_child [F, F, T] and gross earned [0, 500, 0].
  • sc_tanf_dependent_child.yaml:133 Case 7 lists minor_parent first and expects [F, F, T] and [500, 0, 0].
  • integration.yaml:838 Case 28 and integration.yaml:866 Case 29 are the same households, adding countable 250 and sc_tanf 347.
• By hand: need ⌊(15,650 + 2 × 5,500) / 12 × 0.5 = 1,110.42⌋ = 1,110. Countable is 500 less the 50% January disregard, 250. Payment ⌊(1,110 − 250) × 0.4046 = 347.96⌋ = 347. With her wages dropped, ⌊1,110 × 0.4046 = 449.11⌋ = 449.
• Defect and fix, from a Simulation probe of the same household:

	sibling listed first	earner listed first
head 71e1d42269	dependent [F, T, T], gross [0, 0, 0], countable 0, sc_tanf 449	dependent [F, T, T], gross [500, 0, 0], countable 250, sc_tanf 347
fix	dependent [F, F, T], gross [0, 500, 0], countable 250, sc_tanf 347	dependent [F, F, T], gross [500, 0, 0], countable 250, sc_tanf 347

At the head, all four new cases fail (2 of 7 in sc_tanf_dependent_child.yaml, 2 of 29 in integration.yaml). The earner-first cases fail too, because they assert the sibling is no dependent child. With the fix, all pass.

• Mutants: each ran in place in its own process against sc_tanf_dependent_child.yaml and integration.yaml, and was restored with git checkout (git status clean after each).

Mutant	Result	Killed by
(b) heads_household = head_or_spouse	killed	dependent_child 5, 6, 7; integration 26, 28, 29
(c) no_adult & True	killed	dependent_child 2, 5, 6, 7; integration 26, 28, 29
(d) return meets_age_test	killed	dependent_child 2, 5, 6, 7; integration 26, 27, 28, 29
(e) oldest = person.get_rank(person.spm_unit, age) == 0	killed	dependent_child 2, 5, 6, 7; integration 26, 28, 29
(g, new) oldest = person.get_rank(person.spm_unit, -age) == 0, the round 2 head's line	killed	dependent_child 6, 7; integration 28, 29

3. Body (read back after the PATCH, matches the file I sent)

• Mapping: the sentence is now "In a group with no adult, the oldest member heads the household, and members who tie for oldest all head it."
• Tests: the CI sentence now names run 41834176912 at c671a313e3 and its state at 60 minutes (below).
• Overlap: Add South Carolina LIHEAP regular heating assistance #10616 is still at 85d8103fbd. Its candidates.yaml diff against main still adds exactly the six ids us-sc:liheap-energy-burden-test, -income-floor-vs-smi-limit, -low-income-tier-base, -minor-income, -mixed-status-household and -plan-income-checklist. I put in the brief's two sentences.
• impact: and axiom: lines are untouched, left for the hub.

4. Runs (one file per process, after the merge)

• YAML, 15 SC TANF files, 135 cases, all passing (131 + 4 new):
  • integration 29, gross_income_eligible 21, gross_unearned_income 18, sc_tanf 16, need_standard 12, dependent_child 7;
  • countable_earned_income_person, countable_income_eligible, eligible, income_eligible, resources_eligible and ssi_recipient 4 each;
  • benefit_group_size and countable_resources 3 each, gross_earned_income 2.
• Python:
  • test_sc_tanf_child_earnings_support_invariants.py: 4 passed;
  • test_sc_tanf_unearned_income_invariants.py: 2 passed;
  • test_source_conflicts.py: 103 passed.
  • Both invariant files need the dev extra (Hypothesis). Without it the first errors at collection and the second shows 1 passed.
• ruff format --check and ruff check on sc_tanf_dependent_child.py: clean.
• CI: Pull request run 41834176912 at c671a313e3 was still in_progress at 13:59Z, 59 minutes after the 13:00Z push. The jobs API listed 36 jobs: 34 success; Contrib other-shard-1 and other-shard-3 still running. No conclusion yet.

Judgment calls

• The merge. I merged main because 8 commits is more than "a few". It merged cleanly.
• The overlap bullet. The brief replaces its "last two sentences". The bullet's first sentence also ran on into the old entry list ("...only: both add a us-sc: entry at the same place, after the last us-ri: entry (...)"). I kept its opening clause, "It merges cleanly into main, but conflicts with this PR in source_conflicts/candidates.yaml only.", and replaced everything after it with the brief's two sentences. That way the bullet still says where the conflict is.
• The push. I pushed once, at 13:00Z, before the local runs and mutants finished, because the code was final. Nothing that followed needed another push.

Not checked

• I ran no microsimulation and no full suites, and I had no private data. Whether every member of the dataset's no-adult SPM units has a supplied HEAD role remains the hub's claim.
• I read no SC DSS source in this round, so I didn't verify the quotations in the body here.

No follow-up issue proposed.

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

Round 3 review at c671a313e369692e79cbe90a3985408ad4327578; the head hasn't moved and main hasn't moved past the merged 25c4253483.

So far:

• The fix commit b7075ee067 carries round 2's code verbatim (sc_tanf_dependent_child.py:36-43), and its documentation follows the brief's text.
• Ties: a Simulation probe at the head gives 347 for all six orders of a two-way tie (17, 17, baby) and 440 for a three-way tie (16, 16, 16, baby) with the earner first, second, third or last. With a parent aged 35 in the group, the head-or-spouse rule decides in both orders (541).
• Hand values for Cases 6-7 and 28-29 match the YAML.
• CI: Pull request run 41834176912 is the only run at the head, success, 36 of 36 jobs.
• Body (P3s so far): the CI sentence still says the run was in progress; the Tests table still gives 131 cases (integration.yaml 27, sc_tanf_dependent_child.yaml 5), while the head has 135 (29 and 7); the Include farm and partnership earnings in SSI and correct loss netting #10532 bullet says it merges cleanly, but since the main merge it conflicts in ca_state_supplement.yaml, as it does with main.
• Add South Carolina LIHEAP regular heating assistance #10616 (85d8103fbd) still conflicts only in candidates.yaml.

Still running: the test files and mutants.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: REQUEST_CHANGES
c671a31
METHODOLOGY_CHOICE: yes

Independent round 3 review. The head hasn't moved past c671a313e3, and main hasn't moved past the merged 25c4253483, so git rev-list HEAD..origin/main is empty. Since round 2's 71e1d42269 the branch adds two commits:

• b7075ee067 changes sc_tanf_dependent_child.py and adds four YAML cases;
• c671a313e3 merges main. Its SC-relevant change is only main's own candidates.yaml entries.

METHODOLOGY_CHOICE stays yes for the readings round 2 listed: tax-unit head or spouse for "head-of-household", one member's support off the whole group's income, and the regulation above eight persons. Add the hub's no-adult rule, which now makes members who tie for oldest all head the group.

The round 2 P2 is fixed, and the code is correct in every tie I tried. What's left is three stale statements in the body and one test gap. None needs a code change.

Findings

P3. The body's CI sentence is stale again

PR body, "Tests", last paragraph. It says run 41834176912 "was still in progress 60 minutes after the push", with 34 of 36 jobs passed. The run has since finished. It's the only run at this head.

gh api "repos/PolicyEngine/policyengine-us/actions/runs?head_sha=c671a313e3...&per_page=20"
41834176912  Pull request  pull_request  completed  success  13:00:30Z  14:07:25Z
gh api repos/PolicyEngine/policyengine-us/actions/runs/41834176912/jobs?per_page=100
total_count 36; all 36 conclusion "success" (Contrib other-shard-3 last, 14:07:24Z)

Replace the sentence with: "CI: Pull request run 41834176912 at c671a313e3 concluded success; all 36 jobs passed."

P3. The body's Tests section and mutant table describe round 2's head

PR body, "Tests" and "Mutants".

• "Each file ran in its own process at 71e1d42269. Every SC TANF YAML file passes, 131 cases in all" is stale. The table gives integration.yaml 27 and sc_tanf_dependent_child.yaml 5. At this head the files hold 29 and 7 cases (grep -c '^- name:'), 135 in all, and all pass (below).
• "This round added..." names only Cases 5, 26 and 27. The mutant table's rows 8-10 cite only the older cases, and it has no row for the round 2 head's get_rank line.

Replace:

• "Each file ran in its own process at 71e1d42269. Every SC TANF YAML file passes, 131 cases in all:" with "Each file ran in its own process at c671a313e3. Every SC TANF YAML file passes, 135 cases in all:".
• In the table, integration.yaml | 27 with 29, and sc_tanf_dependent_child.yaml | 5 with 7.
• After the "This round added..." sentence, add: "Round 3 added sc_tanf_dependent_child.yaml Cases 6 and 7 and integration.yaml Cases 28 and 29: two 17-year-olds, one earning, and a baby, with no adult and no role inputs, listed in two orders."
• In the mutant table, rows 8, 9 and 10's "Killed by" cells with, in order, "sc_tanf_dependent_child.yaml Cases 5, 6, 7; integration.yaml Cases 26, 28, 29", "sc_tanf_dependent_child.yaml Cases 2, 5, 6, 7; integration.yaml Cases 26, 28, 29" and "sc_tanf_dependent_child.yaml Cases 2, 5, 6, 7; integration.yaml Cases 26-29".
• Add a row: "| 11 | Pick one oldest member by rank (oldest = person.get_rank(person.spm_unit, -age) == 0) | killed | sc_tanf_dependent_child.yaml Cases 6, 7; integration.yaml Cases 28, 29 |".

P3. The #10532 overlap bullet is false since the main merge

PR body, "Overlap": "#10532. It merges cleanly with this PR." #10532 (4cab6b38ae) now conflicts with this head in policyengine_us/tests/policy/baseline/gov/states/ca/cdss/state_supplement/ca_state_supplement.yaml, three hunks. It conflicts with main in the same file, so the conflict came in with the main merge, not with this PR's own changes.

git merge-tree --write-tree --name-only origin/main pr-10532   -> ca_state_supplement.yaml
git merge-tree --write-tree --name-only pr-10581 pr-10532      -> ca_state_supplement.yaml

Replace the bullet with: "#10532. It conflicts with main in ca_state_supplement.yaml, and with this PR in the same file only, because this PR merged main; this PR adds no conflict of its own."

The #10527 bullet's file list is short too. It now also conflicts in nj_eitc.py and nj_eitc.yaml, identically against main and against this PR. Its "Against this PR the conflicts are identical" still holds. To keep it exact, replace "in Makefile, me_eitc.py and a us-me: hunk of candidates.yaml" with "in Makefile, me_eitc.py, nj_eitc.py, nj_eitc.yaml and a us-me: hunk of candidates.yaml".

P3. No test pins "oldest" against a younger minor near the oldest's age

policyengine_us/variables/gov/states/sc/tanf/sc_tanf_dependent_child.py:42. My mutant (h), oldest = age >= person.spm_unit.max(age) - 1, survives both YAML files. It makes every member within a year of the oldest a head. It survives because each no-adult case pairs 17-year-olds with a baby aged 0. Any widening up to max(age) - 16 survives the same way. Witness: sibling 17, minor parent 16 earning $6,000 a year, baby 0, 2026-01, no role inputs. Under the hub's "oldest member" rule only the sibling heads, so the 16-year-old is a dependent child and her wages drop out: need 1,110, countable 0, ⌊1,110 × 0.4046 = 449.11⌋ = 449. Under (h) she heads too, and her $250 counts, giving 347.

Add this case to sc_tanf_dependent_child.yaml, and the same household with these four outputs to integration.yaml:

- name: Case 8, a younger minor in a group with no adult is a dependent child.
  period: 2026-01
  input:
    people:
      sibling:
        age: 17
      minor_parent:
        age: 16
        employment_income_before_lsr: 6_000
      baby:
        age: 0
    households:
      household:
        members: [sibling, minor_parent, baby]
        state_code: SC
  output:
    # No tax-unit role is input. The 17-year-old is the oldest member of a
    # group with no adult, so she alone heads the household; the 16-year-old
    # is a dependent child and Section 7.5 excludes her earnings.
    sc_tanf_dependent_child: [false, true, true]
    sc_tanf_gross_earned_income: [0, 0, 0]
    sc_tanf_countable_income: 0
    # floor(1,110 x 0.4046 = 449.11) = 449.
    sc_tanf: 449

I ran that case as a separate file. It passes at this head (1 passed). Under (h) it fails: sc_tanf_dependent_child@2026-01: [0. 0. 1.] differs from [0. 1. 1.]. I restored the source afterwards, and git status was clean. If the hub reads "oldest" differently for near ties, this case is the place to record that reading.

Checked and sound

• The P2 fix. sc_tanf_dependent_child.py:36-43 is round 2's code verbatim (git show b7075ee067). oldest = age == person.spm_unit.max(age) replaces the get_rank line. The documentation (lines 14-20) carries the brief's text, and the body's Mapping sentence carries the replacement round 2 asked for. ruff format --check and ruff check pass on the file.

• Ties, by Simulation probe at this head, 2026-01, no role inputs:

  • Two-way tie (17, 17, baby 0), the earner in all six orders, including first, middle and last: the 17-year-olds are never dependent children, the baby always is, countable 250, sc_tanf 347 every time.
  • Three-way tie (16, 16, 16, baby 0), the earner first, second or third, and last after the baby: all three 16-year-olds head, countable 250. Size 4 need is ⌊(15,650 + 3 × 5,500) / 12 × 0.5 = 1,339.58⌋ = 1,339, so ⌊1,089 × 0.4046 = 440.61⌋ = 440, the probe's value in every order.
  • With an adult: parent 35, two 17-year-olds (one earning) and a baby, with the parent listed first or last. The head-or-spouse rule decides. Only the parent heads, the 17-year-olds are dependent children, countable 0, and ⌊1,339 × 0.4046 = 541.76⌋ = 541 in both orders.
  • Adult 20 input as head and an earning 17-year-old input as spouse, with a baby: the 17-year-old counts as a parent, 347.
  • Fractional ages (17.0 sibling, 17.5 earner): only the strictly oldest heads, 347.

• Hand values, computed from the parameters before reading the YAML. At 2026-01 the model reads the guideline at 2025-10-01: $15,650 + $5,500. Need rate 0.5, payment rate 0.4046, a 50% disregard for 4 months then $100, and the minimum payment $10. Size 3 need ⌊26,650 / 12 × 0.5 = 1,110.42⌋ = 1,110. The minor parent's 6,000 / 12 = 500 less 50% leaves 250. The gross limit ⌊1.85 × 1,110 = 2,053.5⌋ = 2,053 exceeds 500. The payment is ⌊(1,110 − 250) × 0.4046 = 347.96⌋ = 347, or ⌊449.11⌋ = 449 with her wages dropped. Cases 6 and 7 (sc_tanf_dependent_child.yaml) and 28 and 29 (integration.yaml) assert exactly these: roles [F, F, T], gross [0, 500, 0] or [500, 0, 0], countable 250, sc_tanf 347. Each pair lists the two orders, so the tests assert order independence directly, not a weaker claim.

• The defects. At origin/main (worktree, same probe):

  • parent 35, child 16 earning $6,000 a year, child 8: countable 250, sc_tanf 348.12;
  • the tie household gives 348.12 in both orders.

  At this head the first gives countable 0 and 449, as the PR intends, and the tie gives 347 in both orders. Round 2's order dependence doesn't exist on main. It was this PR's, and it's gone.

• Tests, one file per process at this head:

  • YAML: the 15 SC TANF files, 135 cases, all passing. integration 29, gross_income_eligible 21, gross_unearned_income 18, sc_tanf 16, need_standard 12, dependent_child 7, countable_earned_income_person 4, countable_income_eligible 4, eligible 4, income_eligible 4, resources_eligible 4, ssi_recipient 4, benefit_group_size 3, countable_resources 3, gross_earned_income 2.
  • Python: test_sc_tanf_child_earnings_support_invariants.py 4 passed, test_sc_tanf_unearned_income_invariants.py 2 passed, code_health/test_source_conflicts.py 103 passed. These match the hand-off.

• Mutants. Each ran in place in its own process, one process per test file, against sc_tanf_dependent_child.yaml and integration.yaml. Each was restored with git checkout, and git status was clean after each.

Mutant	Result	Killed by
(b) heads_household = head_or_spouse	killed	dependent_child 5, 6, 7; integration 26, 28, 29
(c) head_or_spouse | (no_adult & True)	killed	dependent_child 2, 5, 6, 7; integration 26, 28, 29
(d) return meets_age_test	killed	dependent_child 2, 5, 6, 7; integration 26, 27, 28, 29
(e) oldest = person.get_rank(person.spm_unit, age) == 0	killed	dependent_child 2, 5, 6, 7; integration 26, 28, 29
(g) oldest = person.get_rank(person.spm_unit, -age) == 0 (round 2's line)	killed	dependent_child 6, 7; integration 28, 29
(h, mine) oldest = age >= person.spm_unit.max(age) - 1	survived	none; the proposed Case 8 kills it (P3 above)
(i, mine) oldest = age == person.spm_unit.min(age)	killed	dependent_child 2, 5, 6, 7; integration 26, 28, 29

Rows (b)-(e) and (g) match the hand-off's table exactly.

• Readers. sc_tanf_dependent_child is read only by sc_tanf_gross_earned_income and sc_tanf_child_support_deduction (grep of policyengine_us/ outside tests). The tie change moves a tied, non-parent member from dependent child to head. That member's earnings now count, as they do on main, and any child support they pay is now deducted. Neither reader is made worse than main. The rest of the reader chain is unchanged since round 2's list.
• Baseline movement. The tie change moves only no-adult groups with a tie for oldest and no input roles. Per the hub's dataset reading that's household calculations only, which I couldn't check (below). The PR as a whole moves SC TANF at baseline, as rounds 1 and 2 said. The impact: and axiom: TheAxiomFoundation/rulespec-us#1728 queued lines both exist.
• Overlap.
  • Add South Carolina LIHEAP regular heating assistance #10616 (85d8103fbd, unchanged) merges cleanly into main. Against this head git merge-tree conflicts in source_conflicts/candidates.yaml only. Its diff against main adds exactly the six ids us-sc:liheap-energy-burden-test, -income-floor-vs-smi-limit, -low-income-tier-base, -minor-income, -mixed-status-household and -plan-income-checklist. Both files are sorted by id (this head 359 entries, Add South Carolina LIHEAP regular heating assistance #10616 363), and in both the us-sc: block sits between us-ri:ssp-title-xix-couple-amount and us-sd:liheap-2025-published-income-limit. Since liheap < tanf, the sorted order is the six LIHEAP entries, then us-sc:tanf-need-standard-above-eight, as the body's bullet says.
  • Open-PR searches on each changed variable name return only this PR. A search on candidates.yaml returns Follow state statutes where forms differ: NM medical exemption, VT CDCC, NY/NYC household credit, PR EITC #10300, Add Colorado LEAP regular heating assistance #10609, Compute the California state supplement per person #10601, Reduce Colorado's federal deduction addback by the state income tax already added back #10094, Add Delaware LIHEAP regular heating assistance #10612 and Follow the statute for the NYC school tax rate reduction, IN supplemental homestead, VA 529 and WA senior/disabled property tax exemption #10330, which all merge cleanly with this head. Round summed resources to whole cents before exact-limit asset tests #10639 does too.
  • Include farm and partnership earnings in SSI and correct loss netting #10532 and Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 are under the overlap P3 above.
• Scope. git diff --name-only origin/main HEAD touches nothing under .github/ or policyengine_us/tests/policy/baseline/partners/. The changelog fragment is a top-level changelog.d/sc-tanf-child-earnings-support-whole-dollars.fixed.md.
• CI. Run 41834176912 is the latest, and only, run at the head: success, 36 of 36 jobs success.

What I couldn't check

• Sources. From this session every source host returns a proxy 403: dss.sc.gov, scstatehouse.gov, web.archive.org, govinfo.gov, ecfr.gov, uscode.house.gov, law.cornell.edu and ssa.gov. So I couldn't confirm any quotation in the body from Policy Manual Vol. 67 §§4.10, 7.5, 8.11, 8.15 and 8.16 or S.C. Code Regs. 114-1140. The tie rule rests on the hub's ruling and round 2's reasoning, not on a source I read.
• Microdata. I couldn't read the default dataset, so the claim that its no-adult SPM units carry supplied HEAD roles, which keeps the tie change to household calculations, is unverified here.
• Population impact. I ran no microsimulation.
• rulespec-us#1728. That repository is outside this session's GitHub scope.

---

Generated by Claude Code

[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: 39ddd6888732e686b2e69d5be68f6be79bc889ab

One commit on top of c671a313e3: 39ddd68887 adds YAML cases only. No one else pushed to the branch while I worked. Main has two commits since this branch's last merge of main (5f2edf8f6d #10601, 0315452583). git merge-tree of this head with origin/main is clean and no check needed main, so I didn't merge it.

What the round 3 review asked for, and what I did

1. A younger minor near the oldest's age (fourth P3).
   • policyengine_us/tests/policy/baseline/gov/states/sc/tanf/sc_tanf_dependent_child.yaml:152-176: Case 8, "Case 8, a younger minor in a group with no adult counts as a dependent child." It uses the review's household and four outputs, with the brief's first comment.
   • policyengine_us/tests/policy/baseline/gov/states/sc/tanf/integration.yaml:887-911: Case 30, the same household, outputs and comments. Its name is the same with "Case 30".
   • Hand check, from the review's parameters: size 3 need ⌊(15,650 + 2 × 5,500) / 12 × 0.5 = 1,110.42⌋ = 1,110. The 16-year-old is a dependent child, so countable income is 0 and the payment is ⌊1,110 × 0.4046 = 449.11⌋ = 449.
   • Both pass at this head. Under mutant (h), oldest = age >= person.spm_unit.max(age) - 1, both fail: sc_tanf_dependent_child.yaml Case 8 and integration.yaml Case 30, and nothing else.
2. The body (first three P3s).
   • Tests: the sentence now names 39ddd68887 and 137 cases. The table gives integration.yaml 30 and sc_tanf_dependent_child.yaml 8.
   • I replaced the "This round added" sentence with the brief's text.
   • Mutants: I reran rows 8–10 at this head and filled in their cells. I added rows 11 and 12. The mutants paragraph now says "Row 2 ran at 71e1d42269 … Rows 8–12 ran at 39ddd68887", and "The property tests killed none of row 2 or rows 8–12". Without those edits, both sentences would have been false.
   • Overlap: git merge-tree --write-tree --name-only gives the same files as the review (below). I replaced the Include farm and partnership earnings in SSI and correct loss netting #10532 bullet and the Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 file list with the brief's text. I also reran Add South Carolina LIHEAP regular heating assistance #10616 because the section's lead sentence claims this head. The bullet still holds: it is clean against main and conflicts with this head only in source_conflicts/candidates.yaml.
   • I left the impact: and axiom: lines as they were. I read the body back after the PATCH, and it is byte-identical to what I sent.
3. Runs: see below.

Tests at 39ddd68887

Each YAML file ran in its own process, at most 4 at a time. All 15 SC TANF files pass, 137 cases:
integration 30, gross_income_eligible 21, gross_unearned_income 18, sc_tanf 16, need_standard 12, dependent_child 8, countable_earned_income_person 4, countable_income_eligible 4, eligible 4, income_eligible 4, resources_eligible 4, ssi_recipient 4, benefit_group_size 3, countable_resources 3, gross_earned_income 2.

test_sc_tanf_child_earnings_support_invariants.py passed 4 in every mutant run below. No Python file changed, so ruff format --check and ruff check had nothing to run on. No microsimulation was run.

Mutants

Each mutant ran in place on sc_tanf_dependent_child.py. For each one, the 7 YAML files the body names ran one process per file, then the property file ran. I restored the file with git checkout after each, and git status showed only my uncommitted YAML at the time.

#	Mutant	Killed by
8	heads_household = head_or_spouse	dependent_child 5, 6, 7, 8; integration 26, 28, 29, 30
9	oldest = True	dependent_child 2, 5, 6, 7, 8; integration 26, 28, 29, 30
10	return meets_age_test	dependent_child 2, 5, 6, 7, 8; integration 26–30
11	oldest = person.get_rank(person.spm_unit, -age) == 0	dependent_child 6, 7; integration 28, 29
12	oldest = age >= person.spm_unit.max(age) - 1	dependent_child 8; integration 30

The other five YAML files passed under every mutant, and the property file passed 4 every time. None survives.

Overlap (git merge-tree --write-tree --name-only)

• Include farm and partnership earnings in SSI and correct loss netting #10532 (4cab6b38ae): it conflicts in ca_state_supplement.yaml against origin/main, against this head, and against 25c4253483, the main this PR merged. This PR's diff against main doesn't touch that file. Main's new Compute the California state supplement per person #10601 also edits the file, so the conflict hunks against current main differ from those against this head. The file list is the same.
• Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 (5a92f6cc0c, draft): against both main and this head it conflicts in Makefile, me_eitc.py, nj_eitc.py, nj_eitc.yaml and candidates.yaml. The conflict text is identical once the marker labels are normalized. The only candidates.yaml hunk is the us-me: one. This PR's us-sc:tanf-need-standard-above-eight merges cleanly.
• Add South Carolina LIHEAP regular heating assistance #10616 (85d8103fbd): clean against main, and against this head it conflicts in candidates.yaml only.

CI

Pull request run 41945984219 at 39ddd68887 hadn't finished 60 minutes after the 01:04Z push. At 02:06Z, 34 of 36 jobs had concluded success. Contrib tests / Full Suite - Contrib (other-shard-1) and (other-shard-3), both started at 01:37Z, were still running. None had failed or been cancelled.

The body's CI sentence still describes run 41834176912 at c671a313e3 as in progress, so it is stale. Item 8 applies once a run finishes, and this one hadn't, so I left that sentence for the hub. When the run passes, it should read: "Pull request run 41945984219 passed all 36 jobs at 39ddd68887."

Judgment calls

• The brief didn't name Case 30, so I gave it Case 8's name with the number changed.
• I didn't apply the review's CI-sentence P3, because the run at this head hadn't finished when I handed off (see CI).
• I edited two more mutants-paragraph sentences, which would otherwise have named only 71e1d42269 and rows 8–10 (see item 2).
• No primary source was fetched this round; the change quotes nothing new.

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

Round 4 delta review at 39ddd6888732e686b2e69d5be68f6be79bc889ab (unchanged since the round 4 hand-off).

So far:

• The only commit since c671a313e3 is 39ddd68887, which adds sc_tanf_dependent_child.yaml Case 8 and integration.yaml Case 30 (YAML only).
• All 15 SC TANF YAML files pass at the head, each in its own process: 137 cases, matching grep -c '^- name:' and the body's table row by row.
• Hand values for Cases 8 and 30 (2026-01, guideline $15,650 + $5,500 read at 2025-10-01): need ⌊26,650 / 12 × 0.5 = 1,110.42⌋ = 1,110; with the 16-year-old's wages excluded, countable 0 and ⌊1,110 × 0.4046 = 449.11⌋ = 449. Both cases assert these values.
• Pull request run 41945984219 at the head concluded success, 36 of 36 jobs.
• No source host is reachable from this session (proxy 403 or DNS failure for dss.sc.gov, scstatehouse.gov, aspe.hhs.gov, govinfo.gov, ecfr.gov, federalregister.gov and web.archive.org).

Still to do: mutants 8-12 plus one of my own, the overlap merge-trees, main, readers and the body.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: APPROVE
39ddd68
METHODOLOGY_CHOICE: yes

Independent round 4 delta review. The head hasn't moved past 39ddd68887. The one commit since round 3's c671a313e3 adds sc_tanf_dependent_child.yaml Case 8 and integration.yaml Case 30 and changes no code. The PR does everything items 1–3 of the brief asked: Case 8 and Case 30 have the brief's name and comment, the body edits are in place, and every file has been run. Every body claim I could check holds, and I found no wrong result. Both findings are P3s, and both come from one rule interaction that no test reaches. The hub should rule on it, but nothing in this round's scope blocks the PR.

METHODOLOGY_CHOICE stays yes for the readings earlier rounds listed: tax-unit head or spouse for "head-of-household", the oldest member heading a group with no adult (with ties all heading), one member's support coming off the whole group's income, and the regulation above eight persons. P3-1 names one more case of the oldest-member reading that the hub hasn't ruled on explicitly.

Findings

P3-1. No test pins how the oldest-member rule combines with input roles

policyengine_us/variables/gov/states/sc/tanf/sc_tanf_dependent_child.py:41-43. no_adult is computed from age only, so in a group with no adult, the oldest member heads even when another member holds the tax-unit head role. That matches the variable's documentation ("The tax unit's head or spouse heads it, and in a group with no adult the oldest member does"). But no test reaches the case, so this mutant survives every SC TANF file and the property file:

no_adult = ~person.spm_unit.any(head_or_spouse)        # m2

Witness, 2026-01, roles input. A 16-year-old minor parent is input as tax-unit head. Her 17-year-old sibling earns $6,000 a year and is input as her tax dependent, and her baby is too. All three are in one SPM unit in SC.

head:  is_tax_unit_head_or_spouse [T, F, F]  sc_tanf_dependent_child [F, F, T]  gross_earned [0, 500, 0]  countable 250  sc_tanf 347
m2:    is_tax_unit_head_or_spouse [T, F, F]  sc_tanf_dependent_child [F, T, T]  gross_earned [0, 0, 0]    countable 0    sc_tanf 449

Hand values: size 3 need is ⌊26,650 / 12 × 0.5 = 1,110.42⌋ = 1,110. With the sibling heading, ⌊(1,110 − 250) × 0.4046 = 347.96⌋ = 347. With the sibling a dependent child, ⌊1,110 × 0.4046 = 449.11⌋ = 449.

The case is reachable in the microsimulation. The body says the default dataset supplies every role, so in any SPM unit with no adult whose oldest member isn't a tax-unit head or spouse, this clause adds the oldest as a second head. Two readings are possible: "the oldest heads only when no member holds a head role" (m2), or "both head" (the head). The sources I could read don't settle it (see "What I couldn't check"). Ask the hub to rule, then pin the ruling. If the hub keeps the head's reading, append this to sc_tanf_dependent_child.yaml, and the same household with the same outputs to integration.yaml:

- name: Case 9, in a group with no adult the oldest member heads it even when a younger member is input as the tax unit head.
  period: 2026-01
  input:
    people:
      minor_parent:
        age: 16
        is_tax_unit_head: true
      sibling:
        age: 17
        is_tax_unit_dependent: true
        employment_income_before_lsr: 6_000
      baby:
        age: 0
        is_tax_unit_dependent: true
    tax_units:
      tax_unit:
        members: [minor_parent, sibling, baby]
    spm_units:
      spm_unit:
        members: [minor_parent, sibling, baby]
    households:
      household:
        members: [minor_parent, sibling, baby]
        state_code: SC
  output:
    # The 16-year-old heads as the tax unit head, and the 17-year-old heads
    # as the oldest member of a group with no adult, so his earnings count.
    sc_tanf_dependent_child: [false, false, true]
    sc_tanf_gross_earned_income: [0, 500, 0]
    # 6,000 / 12 = 500 a month, less the 50% January disregard: 250.
    sc_tanf_countable_income: 250
    # floor((1,110 - 250) x 0.4046 = 347.96) = 347.
    sc_tanf: 347

If the hub picks the m2 reading instead, change line 41 to the m2 line, and change the case's expected values to [false, true, true], [0, 0, 0], 0 and 449.

P3-2. The body suggests the no-adult rule doesn't act on the microsimulation

PR body, "Rules", item 1, last sentence of the Mapping paragraph: "The default dataset supplies every person's tax-unit role, so these limits apply to household calculations, not the microsimulation." That holds for the role-inference limits. But the sentence just before it states the no-adult oldest-member rule, and a reader measuring impact would take that rule to be household-only too. It isn't: the P3-1 witness gives a household with supplied roles where the rule changes the payment (347 vs 449). I can't see the dataset, so I can't say how many SPM units it reaches. Replace the sentence with:

The default dataset supplies every person's tax-unit role, so the role-inference limits apply to household calculations only. The no-adult rule still acts in the microsimulation: in a group with no adult, the oldest member heads it even when another member is the tax unit's head.

Checked and sound

• Head, main and CI. pulls/10581 gives head 39ddd6888732e686b2e69d5be68f6be79bc889ab, and its last commit is dated 01:04:30Z. origin/main is 0315452583, two commits past this PR's merged main: 5f2edf8f6d (Compute the California state supplement per person #10601, the California state supplement) and 0315452583 (version bump). git merge-tree --write-tree --name-only origin/main pr-10581 is clean, and neither commit touches SC TANF, the poverty guideline or role inference (git diff --name-only pr-10581...origin/main | grep -iE 'sc/|tanf|fpg|tax_unit_head|is_child' matches nothing). That matches the hand-off. Pull request run 41945984219 at the head is completed / success, with 36 of 36 jobs success, so the body's CI sentence is correct.
• The defect at main. For the Case 8 household at 2026-01, Simulation at origin/main gives gross earned [0, 500, 0], countable 250 and sc_tanf 348.12: the 16-year-old's wages count, and the payment isn't floored. At the head it gives sc_tanf_dependent_child [F, T, T], gross earned [0, 0, 0], countable 0 and 449.
• Cases 8 and 30, computed by hand before running them. The parameter tree gives a guideline of $15,650 + $5,500 at 2025-10-01, which is what January 2026 reads. It also gives a need rate of 0.5, a gross limit factor of 1.85, a payment rate of 0.4046 (from 2024-10-01), a 50% disregard for 4 months and a $10 minimum.
  • Size 3 need: ⌊1,110.42⌋ = 1,110. Gross limit: ⌊2,053.5⌋ = 2,053.
  • Under the oldest-member rule, only the 17-year-old heads, so the 16-year-old is a dependent child and Section 7.5 (as the body quotes it) excludes her earnings. Countable is 0, and the payment is ⌊1,110 × 0.4046 = 449.106⌋ = 449.
  • Under mutant (h), she heads too. Countable is 500 − 50% = 250, and the payment is ⌊860 × 0.4046 = 347.96⌋ = 347.
  • Both cases assert [F, T, T], [0, 0, 0], 0 and 449. The gross earned output pins the exclusion per person, not just the total, so each case asserts the full claim, not a weaker one.
• Tests, one file per process at the head. All 15 SC TANF YAML files pass, 137 cases. Per file: integration 30, gross_income_eligible 21, gross_unearned_income 18, sc_tanf 16, need_standard 12, dependent_child 8, and countable_earned_income_person, countable_income_eligible, eligible, income_eligible, resources_eligible and ssi_recipient 4 each. benefit_group_size and countable_resources have 3 each, and gross_earned_income 2. Each count equals grep -c '^- name:' and the body's table.
  • Python: test_sc_tanf_child_earnings_support_invariants.py 4 passed, test_sc_tanf_unearned_income_invariants.py 2 passed, code_health/test_source_conflicts.py 103 passed.
  • ruff format --check and ruff check pass on variables/gov/states/sc/tanf/.
• Mutants. Each ran on a scratch worktree at the head, in its own process. Each ran against the seven YAML files the body names, one process per file, and then the property file. The worktree was restored after each, and git status was clean.

Row	Mutant	Killed by	Property file
8	heads_household = head_or_spouse	dependent_child 5, 6, 7, 8; integration 26, 28, 29, 30	4 passed
9	head_or_spouse | (no_adult & True)	dependent_child 2, 5, 6, 7, 8; integration 26, 28, 29, 30	4 passed
10	return meets_age_test	dependent_child 2, 5, 6, 7, 8; integration 26–30	4 passed
11	oldest = person.get_rank(person.spm_unit, -age) == 0	dependent_child 6, 7; integration 28, 29	4 passed
12	oldest = age >= person.spm_unit.max(age) - 1	dependent_child 8; integration 30	4 passed
m1 (mine)	oldest = age >= person.spm_unit.max(age) - 0.5	survived	4 passed
m2 (mine)	no_adult = ~person.spm_unit.any(head_or_spouse)	survived; see P3-1	4 passed

Rows 8–12 match the body's "Killed by" cells exactly. The body's sentence "The property tests killed none of row 2 or rows 8–12" holds for rows 8–12. Row 12's failure message is sc_tanf_dependent_child@2026-01: [0. 0. 1.] differs from [0. 1. 1.].

• m1 survives because it's equivalent at whole-number ages, which is all the YAML and the dataset use. I don't ask for a case.
• Body.
  • The "Tests" sentence names 39ddd68887 and 137 cases, and the table matches.
  • The "This round added" paragraph is the brief's text. It correctly describes Cases 5/26 (a minor parent with no adult and no role inputs), 6–7/28–29, 8/30 and 27 (17-year-old input as is_tax_unit_spouse, with a 20-year-old spouse and a baby).
  • The two edited mutants-paragraph sentences ("Row 2 ran at 71e1d42269 … Rows 8–12 ran at 39ddd68887" and "killed none of row 2 or rows 8–12") are true for rows 8–12 by my rerun. I didn't rerun row 2.
  • The impact: and axiom: TheAxiomFoundation/rulespec-us#1728 queued lines exist.
• Overlap, from a full-history clone. A shallow clone gives spurious conflicts and "unrelated histories" here, so unshallow before rerunning these.
  • Include farm and partnership earnings in SSI and correct loss netting #10532 (4cab6b38ae) conflicts in ca_state_supplement.yaml only, against origin/main, against this head and against 25c4253483, the main this PR merged. This PR's own diff (origin/main...pr-10581) doesn't touch that file. The bullet is correct.
  • Apply the Maryland and Maine EITC waivers of the 32(m) SSN requirement #10527 (5a92f6cc0c, draft) conflicts in Makefile, me_eitc.py, nj_eitc.py, nj_eitc.yaml and candidates.yaml, against both main and this head. Once marker labels are normalized, the conflict text is identical in all five files. The only candidates.yaml hunk holds four us-me: ids, and us-sc:tanf-need-standard-above-eight isn't in it. The bullet is correct.
  • Add South Carolina LIHEAP regular heating assistance #10616 (85d8103fbd) merges cleanly into main. Against this head it conflicts only in candidates.yaml, in one hunk: its six us-sc:liheap-* ids against this PR's us-sc:tanf-need-standard-above-eight. The bullet is correct.
  • Open-PR searches for sc_tanf_dependent_child and sc_tanf return only Exclude SC TANF dependent children's earnings, deduct child support paid, and round to whole dollars #10581.
• Readers. This round changes no code, so no reader changes. Outside tests, sc_tanf_dependent_child is read by sc_tanf_gross_earned_income and sc_tanf_child_support_deduction, and the deduction by sc_tanf_countable_income (grep of policyengine_us/).
• Baseline. This round's commit can't move any output. The PR as a whole moves SC TANF at baseline, as earlier rounds said, and P3-1 names one more path through which the no-adult rule reaches the data.
• Scope. git diff --name-only origin/main...pr-10581 touches nothing under .github/ or policyengine_us/tests/policy/baseline/partners/. A two-dot diff against current main lists a partners file, but that change comes from main's Compute the California state supplement per person #10601, not this PR.

What I couldn't check

• Sources. Every host I tried is blocked from this session: dss.sc.gov, scstatehouse.gov, aspe.hhs.gov, federalregister.gov, govinfo.gov, ecfr.gov and web.archive.org return a proxy 403 or don't resolve. So I couldn't confirm any quotation in the body from Policy Manual Vol. 67 §§4.10, 7.5, 8.11, 8.15 and 8.16 or S.C. Code Regs. 114-1140, nor the 2025 guideline outside the repo's gov/hhs/fpg.yaml. The oldest-member rule, and the P3-1 question, rest on the hub's ruling, not on a source I read.
• Microdata. I can't read the default dataset, so I can't say how many SPM units with no adult have an oldest member who isn't a tax-unit head or spouse (P3-1, P3-2).
• Population impact. I ran no microsimulation.
• Row 2 and rows 1, 3–7 of the mutant table. This round didn't change them, and I didn't rerun them.
• rulespec-us#1728. That repository is outside this session's GitHub scope.

---

Generated by Claude Code

[DTrim99]

PR Review: Exclude SC TANF dependent children's earnings, deduct child support paid, and round to whole dollars

Summary

This PR adds three South Carolina Family Independence (TANF) rules that main lacks. Main's SC TANF code hasn't changed since this branch's merge base: #9687 and #10184 are already in that base. The rules are: (1) dependent children's earned income is excluded; (2) child support paid outside the home is deducted after the earned income disregards; (3) the need standard, gross income limit and payment are rounded down to whole dollars.

I read the governing regulations (S.C. Code Regs. 114-1130, 114-1140 and 114-1150, Chapter 114 PDF) and TANF Policy Manual Vol. 67 (sha256 cac6566b…, the same file the PR cites).

• Sound: rule 1 and the need-standard and payment floors. 114-1140(N)(1) states rule 1 more directly than the manual. All 8 sizes of the §8.15 table match. The above-eight candidate entry is correct.
• Wrong: two results contradict the regulations.
  • The child support deduction pools the whole group's income, but 114-1140(L) deducts support "from the income of the payor".
  • In a group with no adult, the oldest-member rule overrides input tax-unit roles. A 17-year-old input as a dependent then has his wages counted, which 114-1140(N)(1) forbids.
• Other issues:
  • The gross-limit floor follows the manual's "drop the cents" over 114-1150(A)'s "185 percent of the need standard", and the choice isn't recorded as a source conflict.
  • Two new payer conditions are unpinned.

CI is green (38/38). The branch now conflicts with main in source_conflicts/candidates.yaml; the conflict is trivial to resolve.

Critical (must fix)

C1. The child support deduction comes off the whole group's income; 114-1140(L) takes it only from the payer's income.
policyengine_us/variables/gov/states/sc/tanf/income/sc_tanf_child_support_deduction.py:28-39; integration.yaml:657 (Case 22).

S.C. Code Regs. 114-1140(L) (Chapter 114, p.212):

L. Disregard of Child Support Payments Made by a Budget Group Member to a Non-Budget Group Member. Child support payments made by a budget group member to legal dependents who are not budget group members may be deducted from the income of the payor.

114-1130(B)(3) (p.203) treats a stepparent the same way: "The stepparent's income may be reduced by any child support paid to legal dependents outside the home."

The manual can be read the same way.

• §8.16 D (p.134) says "Subtract from the remaining earned income of any adult, the amount of actual child support paid outside the home."
• §8.16 F says "Subtract from the unearned income, child support paid by any adult not deducted from the earned income in Step D."

The PR body reads "any adult" in D and F, and §8.11's "deducted as income to the BG", as pooling. Even on that reading, CONTRIBUTING.md#source-conflicts ("Regulations are law"; "Statute first") makes the regulation govern over the manual. The body's own alternative ("payer-only reading: 246") is the regulation's result.

Witness: Case 22, May 2026. Adult A pays $2,400 a year and has no income. Adult B earns $600 a month, and there is one child.

• Under 114-1140(L), A has no income, so the deduction is 0 and countable income is 600 − 100 = 500.
• The payment is then ⌊(1,110 − 500) × 0.4046 = 246.81⌋ = 246. The PR pins 327.

Fix:

• Compute the deduction per payer and sum it: min_(max_(child_support_expense, 0), max_(sc_tanf_countable_earned_income_person + sc_tanf_gross_unearned_income, 0)) for each benefit group member.
• Change Case 22 to deduction 0, countable 500 and sc_tanf 246, and rename it. Cases 21, 23 and 25 don't change, because in each the payer is the only earner.
• Cite 114-1140(L) in the reference.
• Update the body, the changelog line and the rulespec-us#1728 extension (companion test 7 and the "Adult payer" reading).
• If you still read §8.11 as pooling, record that reading as a regulation-versus-manual candidate in candidates.yaml that follows the regulation.

C2. In a group with no adult, the oldest member heads the household even when the inputs make him another member's dependent, so a dependent child's wages count.
policyengine_us/variables/gov/states/sc/tanf/sc_tanf_dependent_child.py:41-44; sc_tanf_dependent_child.yaml:178 (Case 9); integration.yaml:913 (Case 31).

In Case 9 the 16-year-old is input as tax-unit head. The 17-year-old earner is input as her dependent (is_tax_unit_dependent: true), and nothing makes him a parent or married. The code still makes him a head, because no_adult & oldest is computed from ages alone, and his $500 a month counts.

The law leaves no room for that:

• 114-1140(N)(1) (p.213): "The earned income of dependent children is excluded in the FI gross income limit test and the need and benefit determination."
• A dependent child is "under eighteen years of age or under nineteen years of age and a full-time student in a secondary school" (114-1130(E), p.204). The manual glossary adds "unmarried" (p.321).
• §7.5 (p.105) counts a minor's earnings only for "the minor parent [who] is the head-of-household".
• A caretaker relative can join the group only "When a parent is not in the home" (§3.3 A, p.49), and the baby's parent is in the home.

So the 17-year-old is either a dependent child in the group or outside it, and either way his wages don't count. The oldest-member rule stands in for missing roles; the code comment says it exists because "roles are inferred only for adults". Here the roles are input, and the rule overrides them.

• Expected: countable 0 and ⌊1,110 × 0.4046 = 449.11⌋ = 449.
• The PR pins countable 250 and 347.

Fix: apply the age rule only when no member holds a head or spouse role. This is the body's mutant 13.

no_head = ~person.spm_unit.any(head_or_spouse)
heads_household = head_or_spouse | (no_adult & no_head & oldest)

• Change Cases 9 and 31 to [false, true, true], [0, 0, 0], countable 0 and sc_tanf 449, and update the variable's documentation string and the body's sentence "Where a younger member holds the head role, the oldest member … still heads it too."
• Cases 5–8 and 26–30 have no role inputs and don't change.
• The body says every member of the dataset's no-adult SPM units already holds the head role, so the microsimulation doesn't change either.

Should address

S1. The gross income limit floor follows the manual over 114-1150(A), and the choice isn't recorded.
policyengine_us/variables/gov/states/sc/tanf/eligibility/sc_tanf_gross_income_eligible.py:19. The "one cent above the limit" cases in sc_tanf_gross_income_eligible.yaml pin the floor.

114-1150(A) (p.217): "Gross monthly countable earned and unearned income of the Family Independence benefit group must not exceed 185 percent of the need standard by family size."

114-1140(A)(4) (p.210) rounds only "the standard of need or the payment amount". The body's rule 3 cites (A)(4) for all three floors, but the gross-limit floor rests only on the §8.15 note "multiply the Need Standard by 1.85 and drop the cents". §8.1 (p.121) also says "Do not round at any step of the TANF calculation."

Witness: a parent and child in January 2026, with wages of $19,554 a year ($1,629.50 a month).

• The PR's limit is ⌊881 × 1.85⌋ = 1,629, so the group fails the gross test and sc_tanf is 0.
• Under 114-1150(A) the limit is 1,629.85, so the group passes. Countable income after the 50% disregard is 814.75 < 881, and the payment is ⌊(881 − 814.75) × 0.4046 = 26.80⌋ = 26.

Above eight persons, the PR already follows the regulation over a manual note (us-sc:tanf-need-standard-above-eight). Do the same here:

• Drop the floor and move the boundary cases to straddle the exact limit. For size 2, use 1,629.84 to pass and 1,629.86 to fail, which avoids a float32 tie at 1,629.85.
• Record a candidates.yaml entry (for example us-sc:tanf-gross-income-limit-cents, follows: statute).
• If you read the regulation as silent on rounding, so that the manual fills a gap, keep the floor but record that reading in candidates.yaml (CONTRIBUTING.md#source-conflicts, step 1).

S2. Two new payer conditions are unpinned (mutation rule).
sc_tanf_child_support_deduction.py:28-30.

Only integration Cases 21–23 and 25 set child_support_expense, and each time on an adult who neither receives SSI nor is a dependent child. The property test sets support only on adults and gives nobody SSI. Deleting ~person("sc_tanf_ssi_recipient", period) & or & ~person("sc_tanf_dependent_child", period) from payer fails no test.

Fix: after reworking C1, decide each condition and pin it.

• Under a payer-only formula the SSI condition is redundant, because an SSI recipient's counted income is already 0.
• 114-1140(L) says "a budget group member", so a dependent child who pays support could deduct it from his own unearned income.
• Add a case for each condition you keep, for example an SSI-recipient parent who pays support while the other parent earns, and a 17-year-old with Social Security who pays support.

Optional: the max_(child_support_expense, 0) clamp and the where(deduction > 0, …) branch in sc_tanf_countable_income.py:31 also survive deletion. The latter changes sc_tanf_countable_income, not sc_tanf, for a group with a self-employment loss.

Suggestions

1. Cite the regulations alongside the manual.
   • 114-1140(N)(1) belongs on sc_tanf_gross_earned_income and sc_tanf_dependent_child. It names the gross income limit test explicitly, which settles the PR's choice to exclude a child's earnings there.
   • 114-1130(E) belongs on the age test. It says "secondary school", which fits is_in_secondary_school more closely than the manual's "primary or secondary".
   • Cite 114-1140(L) on the deduction and 114-1150(A) on the gross test.
   • Drop the body's implication that 114-1140(A)(4) covers the gross limit.
2. Case 8 / Case 30 comments. "DSS leaves out the 16-year-old's earnings (Section 7.5)" doesn't fit the household as named. She is the minor parent, and the glossary (p.328) defines a minor parent as "head-of-household of the TANF BG", so DSS would count her earnings. The case pins PE's inference rule, because no role input identifies the parent. Say that in the comment and name the members neutrally (older_minor, younger_minor). Once C2 is fixed, inputting is_tax_unit_head: true for the minor parent gives DSS's result.
3. sc_tanf_countable_income.py:26 says "nothing says to round it". §8.1 (p.121) does speak to it: "Do not round at any step of the TANF calculation." Cite it, worded so the source-conflict ratchet doesn't count it.
4. Python property test cost. CLAUDE.md asks for peak memory for new tests that build simulations. The body gives the time (34 s with startup, 7 s of tests) and the CI group but no memory figure. Add the figure, or say it wasn't available.

Verified correct

• Rule 1. The exclusion matches 114-1140(N)(1), §7.5 and §8.16 B. It applies to the gross test as well as the net test and the benefit, as (N)(1) requires.
  • A dependent child's unearned income still counts (Case 20).
  • Pregnancy no longer makes an adult a dependent child (Cases 3 and 24).
  • The age test matches 114-1130(E) and §4.10 (Case 1 covers 17, 18 in school, 18 out of school and 19 in school).
  • A married minor input as spouse is counted (Case 27), which fits the "unmarried" requirement.
• Rule 3, need standard and payment. Both floors match 114-1140(A)(4) and the §8.15 table at all 8 sizes. The identity ⌊⌊x⌋/2⌋ = ⌊x/2⌋ holds for x ≥ 0. Flooring before the $10 test is equivalent to testing first, because ⌊x⌋ ≥ 10 exactly when x ≥ 10 (114-1140(A)(5)).
• Rule 2, order. The deduction comes after the disregards ("after any other budget calculations", §8.11; §8.16 C then D). It doesn't reach the gross test (Cases 21 and 25 keep gross income at 600), and a group with no income gets no deduction (§8.11(2), Case 23).
• Above-eight candidate. The arithmetic holds:
  • size 12: ⌊76,150/24⌋ = 3,172 = 2,256 + 4 × 229;
  • size 13: ⌊81,650/24⌋ = 3,402 against 3,401;
  • size 18: 4,547 against 4,546;
  • sizes 19 and 24: differences of 2 (4,777/4,775 and 5,922/5,920).
• Code checks.
  • Vectorization, entities and periods are correct: yearly age and roles are read through period.this_year, and yearly child_support_expense is read per month.
  • No state-gated int or enum key is added, so the NH Add New Hampshire state supplementation for 2026 #10156 pattern doesn't apply.
  • No income source list changes, so the Integrate modeled state unemployment insurance into income flows #9558 guard doesn't apply.
  • No partner test is touched.
  • The changelog fragment is top-level and typed fixed.
  • Readers: sc_tanf_dependent_child is read only by sc_tanf_gross_earned_income and the deduction, and the deduction only by sc_tanf_countable_income.

Hand checks (2024 guideline 15,060 + 5,380; 2025 guideline 15,650 + 5,500; 2019 guideline 12,490 + 4,420; 2023 guideline 14,580 + 5,140):

Case	Arithmetic	Pinned
integration 1 (2025-01, size 3)	⌊25,820/24 = 1,075.83⌋ = 1,075; ⌊1,075 × 0.4046 = 434.95⌋	434 ✓
integration 2	⌊(1,075 − 250) × 0.4046 = 333.80⌋	333 ✓
integration 3 (size 2)	need ⌊20,440/24⌋ = 851; limit ⌊851 × 1.85 = 1,574.35⌋ = 1,574 < 2,000	0 ✓
integration 6 (2024-01)	⌊19,720/24 = 821.67⌋ = 821; ⌊821 × 0.3372 = 276.84⌋	276 ✓
integration 7 (size 6)	⌊41,960/24 = 1,748.33⌋ = 1,748; ⌊707.24⌋	707 ✓
integration 9 (2020-01)	⌊16,910/24 = 704.58⌋ = 704; ⌊204 × 0.3372 = 68.79⌋	68 ✓
integration 10	⌊25,750/24⌋ = 1,072; ⌊72 × 0.3372 = 24.28⌋	24 ✓
integration 14	limit ⌊1,075 × 1.85 = 1,988.75⌋ = 1,988; ⌊325 × 0.4046 = 131.50⌋	131 ✓
integration 20 (2026-01)	need ⌊26,650/24 = 1,110.42⌋ = 1,110; ⌊910 × 0.4046 = 368.19⌋	368 ✓
integration 21 (May)	E = 600 − 100 = 500, S = 200 → 300; ⌊810 × 0.4046 = 327.73⌋	327 ✓
integration 23 (May)	E = 150, U = 100, S = 300 → min(300, 250) = 250; countable 0	449 ✓
integration 25 (Jan)	E = 300 − 200 = 100; ⌊1,010 × 0.4046 = 408.65⌋	408 ✓
integration 26 (size 2)	⌊21,150/24 = 881.25⌋ = 881; ⌊631 × 0.4046 = 255.30⌋	255 ✓
sc_tanf 6 / 7 / 8	⌊24.02 × 0.4046 = 9.72⌋ = 9 → 0; ⌊24.80 × 0.4046 = 10.03⌋ = 10; ⌊24.20 × 0.4046 = 9.79⌋ = 9 → 0	0 / 10 / 0 ✓
need standard, size 9	⌊59,650/24 = 2,485.42⌋ = 2,485 = 2,256 + 229	2,485 ✓
gross limit size 7; payment size 8	⌊2,027 × 1.85 = 3,749.95⌋ = 3,749; ⌊2,256 × 0.4046 = 912.78⌋ = 912 (table $3749, $912)	✓

Each case computes as the code intends. Case 22 (327) and Cases 9 and 31 (347) are arithmetically right but legally wrong (C1, C2). The §8.15 gross-limit boundary cases depend on S1.

Axiom line assessment

The line is axiom: TheAxiomFoundation/rulespec-us#1728 queued. The issue is OPEN, labelled pe-parity, and records its blocker: the pinned corpus release predates the TANF manual scope, and a billed signed run needs approval.

Its 2026-11-02 extension comment for this PR (written at cdf464ad90) gives:

• the verbatim law (§§4.10, 7.5, 8.11, 8.15, 8.16 and 114-1140(A)(4)–(5)), with each page named as a corpus row;
• the readings PE takes;
• companion tests mapped to integration Cases 18, 19, 25, 21, 23 and 22 and sc_tanf.yaml Case 8.

All of those values match the head, so the line is valid at this head. One gap: the "Readings" bullet doesn't describe the no-adult oldest-member rule, but the extension says an encoding can take head of household as an input, so that doesn't block it.

Once C1, C2 and S1 are fixed, the issue must follow:

• Companion test 7 becomes $246.
• The "Adult payer" reading gives way to 114-1140(L), quoted.
• Add 114-1140(N)(1) and 114-1150(A).

Also reconcile the issue body's need-standard output and its companion test 11 ("it does not yet drop cents", $851.67) with the extension's floored $851.

CI status

gh pr checks 10581: 38 of 38 pass at 95e9baf6ea; none pending or failed.

Branch status

• Behind main: 21 commits.
• Conflict: git merge-tree --write-tree PolicyEngine/main 95e9baf6ea conflicts in source_conflicts/candidates.yaml. Main's Restrict military survivor tax exemptions to eligible spouses #10398 added us-sc:2021-survivor-offset-surviving-spouse at the same spot (after the last us-ri: entry) as this PR's us-sc:tanf-need-standard-above-eight. Keep both, with main's entry first in id order. This conflict is new since the author's last overlap check.
• Other main changes since the merge base:
  • Main hasn't touched SC TANF.
  • Main's role-inference files changed only in documentation (Add state tax comparison outputs and caller documentation #10310).
  • The ratchet's KNOWN_UNMARKED gained an SC income-tax file that doesn't interact with this PR.
• Sibling PR: Add South Carolina LIHEAP regular heating assistance #10616 (SC LIHEAP) will also conflict in candidates.yaml, as the body notes.

Verdict: REQUEST_CHANGES

[DTrim99]

Requesting changes: 2 critical ((1) the child support deduction comes off the whole group's income, but 114-1140(L) takes it only from the payer's income: Case 22 should pay $246, not $327; (2) in a group with no adult, the oldest-member rule overrides input roles, so a dependent 17-year-old's wages count against 114-1140(N)(1): Case 9 should pay $449, not $347), 2 should (the gross-limit floor follows the manual over 114-1150(A), unrecorded; two payer conditions unpinned). Full report in the comment above.

[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: e4e6861e6a938c08e6587f99c6def41360a63176

I picked the PR up at 95e9baf6ea, the head the brief names. Nobody else pushed while I worked: before my single push at 15:06Z, the PR head was still 95e9baf6ea, an ancestor of mine. The six commits on top:

• 96bd29a40f merges main ba9bfdea75 (item 1);
• a1cc4925ae makes the deduction per payer (C1, S2, suggestion 3);
• 4f58041cb9 applies the oldest-member rule only when no member heads the group (C2, suggestions 1 and 2);
• fafee00445 keeps the gross limit's cents (S1);
• 675ca2d115 adds integration.yaml Case 37, the S1 witness through the payment;
• e4e6861e6a adds a fixed property example that pins the countable income floor (see item 8).

Findings reproduced before any change

At 96bd29a40f, whose SC TANF code is 95e9baf6ea's, a Simulation probe of each witness:

• C1 holds. Case 22's household (May 2026) gave deduction 200, countable 300 and sc_tanf 327, though the payer has no income.
• C2 holds. Case 31's household gave sc_tanf_dependent_child [F, F, T], gross earned [0, 500, 0], countable 250 and sc_tanf 347.
• S1 holds. A parent and child with $19,554 a year of wages (January 2026) gave gross income 1,629.50, sc_tanf_gross_income_eligible false and sc_tanf 0.
• S2 holds. Deleting either payer condition left integration.yaml at 31 of 31 and the property file at 4 of 4, in a scratch worktree. No other file sets child_support_expense.

I fetched both sources and read the cited pages: S.C. Code Regs. Chapter 114 (July 2025 compilation) pp. 203–204, 210, 212–213 and 217, and Manual Vol. 67 (sha256 cac6566b…, effective August 1, 2026) pp. 75, 105, 117, 121, 131, 133–134, 321 and 328. Every quotation of the regulation or the manual, below and in the commits, comes from those pages.

Items

1. Merge main. 96bd29a40f merges origin/main ba9bfdea75 (110 commits ahead) with a merge commit. Only source_conflicts/candidates.yaml conflicted. Main's Restrict military survivor tax exemptions to eligible spouses #10398 and Add South Carolina LIHEAP regular heating assistance #10616 had added us-sc:2021-survivor-offset-surviving-spouse and six us-sc:liheap-* entries where this branch had added us-sc:tanf-need-standard-above-eight. I took main's file and put the branch's entry after us-sc:liheap-plan-income-checklist (now candidates.yaml:9796), so the us-sc: block is in id order.
   • git diff origin/main 96bd29a40f touched only the PR's 17 files.
   • Main's changes near this PR are documentation only (is_tax_unit_dependent.py, tax_unit_role_input.py) plus new KNOWN_UNMARKED lines. Nothing in main touches SC TANF.
   • Main's file already has two out-of-order pairs (us:surviving-spouse-child-gross-income-test before us-ar:…, and two us-ar: ids). I left them; the test doesn't check order.
2. C1. sc_tanf_child_support_deduction.py:28-42 takes each member's max_(child_support_expense, 0) from that member's own sc_tanf_countable_earned_income_person + sc_tanf_gross_unearned_income, floored at zero, and sums it over the group. 114-1140(L) is the first reference (:12) and is quoted at :24-27.
   • Case 22 (integration.yaml:658), renamed "support paid by an adult with no income doesn't come off another adult's earnings". By hand: A's deduction is min(200, 0) = 0; B's 600 − 100 = 500 counts; ⌊(1,110 − 500) × 0.4046 = 246.81⌋ = 246. Main gives 246.97 and the old head 327.
   • Cases 21, 23 and 25 don't move: 327 (deduction 200), 449 (deduction min(300, 250) = 250) and 408 (300 − 200 = 100). In each, the payer holds all the group's income. Case 23's name now says "the payer's income".
   • candidates.yaml:9731 adds us-sc:tanf-child-support-deduction-payer. Sources: 114-1140(L) (regulation, p.212) and §§8.11, 8.16 D and F (manual, p.131). The statute reading quotes (L); the other reading quotes §8.16 D's "from the remaining earned income of any adult", F's "paid by any adult" and §8.11's "as income to the BG". It sets follows: statute, state: ruled, and years from: 2026, federal-fiscal.
   • The changelog line now says the support comes off "that member's own income".
3. C2. sc_tanf_dependent_child.py:49-51 adds no_head = ~person.spm_unit.any(head_or_spouse) and heads_household = head_or_spouse | (no_adult & no_head & oldest). The documentation (:16-23) now says the oldest member heads a group with no adult only "where no member holds either role".
   • Cases 9 (sc_tanf_dependent_child.yaml:182) and 31 (integration.yaml:918) expect [false, true, true], [0, 0, 0], countable 0 and ⌊1,110 × 0.4046 = 449.11⌋ = 449 (main 348.12; the old head 347). Both are renamed "a member input as the tax unit head heads it in place of the oldest member".
   • Cases 5–8 and 26–30 keep every expected value and pass.
   • The body's sentence about a younger input head is replaced (item 10).
4. S1. sc_tanf_gross_income_eligible.py:21 drops the floor: the limit is need × 1.85. 114-1150(A) is the first reference (:11) and is quoted at :18-20.
   • sc_tanf_gross_income_eligible.yaml: Cases 4 and 5 are now 1,574.34 (passes) and 1,574.36 (fails) around 851 × 1.85 = 1,574.35. The table cases for sizes 1–8 sit one cent below and above 1,206.20, 1,629.85, 2,053.50, 2,477.15, 2,900.80, 3,324.45, 3,749.95 and 4,173.60; for size 2 that's 1,629.84 and 1,629.86 (:89). I checked every pair in float32 and float64. In float32, 1,629.85 itself ties the limit, as the review warned.
   • candidates.yaml:9763 adds us-sc:tanf-gross-income-limit-cents: 114-1150(A) (p.217) and 114-1140(A)(4) (p.210), both regulations, against the §8.15 note (manual, p.133). It sets follows: statute, state: ruled, and its note quotes §8.1.
   • New integration.yaml:1133 Case 37 is the review's witness. It passes at 1,629.85, and ⌊(881 − 814.75) × 0.4046 = 26.80⌋ = 26 (main 26.91; the old head 0).
   • Stale gross-limit comments are fixed in integration Cases 3 and 14 and in sc_tanf_income_eligible.yaml, whose 851.67 and 1,575.58 predate the need floor.
5. S2. I removed both payer conditions:
   • SSI. Both income variables multiply by ~sc_tanf_ssi_recipient, so an SSI recipient's counted income is 0 and the cap already gives a deduction of 0. The condition was an equivalent mutant (row 16). integration.yaml:955 Case 32 pins the result. An SSI parent pays $2,400 a year while the other parent earns $600 a month: deduction 0, size 2, need 881, ⌊(881 − 500) × 0.4046 = 154.15⌋ = 154 (main 154.25).
   • Dependent child. 114-1140(L) covers "a budget group member", which a dependent child is; the manual's "any adult" belongs to the reading the C1 ruling sets aside. integration.yaml:1001 Case 33: a 17-year-old dependent with $2,400 a year of survivor benefits pays $1,200 a year. The deduction is 100, countable income 100, and ⌊1,010 × 0.4046 = 408.65⌋ = 408 (main 368.35; with an adults-only payer rule, 368). It kills the restored condition (row 17).
   • Clamp, kept. integration.yaml:1038 Case 34 has a support input of −$2,400 a year: deduction 0 and 246. Without the clamp, countable income is 700 and the payment 165. The repo already pins negative expense inputs (NH FAP fuel oil, WY LIEAP heating, CA LIHEAP bottled gas).
   • Branch, replaced (judgment call 1). Cases 35 (:1070) and 36 (:1098) and a fixed property example pin the replacement.
6. Suggestions.

   1. Citations: 114-1140(N)(1) on sc_tanf_gross_earned_income.py:15,21-24 and sc_tanf_dependent_child.py:12,36; 114-1130(E) on the age test (sc_tanf_dependent_child.py:12,29-31); 114-1140(L) and 114-1150(A) as in items 2 and 4. The body no longer cites 114-1140(A)(4) for the gross limit.

   2. Cases 8 and 30: the members are now older_minor and younger_minor. The comments say:

      • no input says which minor is the baby's parent, so the case pins PE's inference rule;
      • DSS counts a head-of-household minor parent's earnings (glossary p.328, §7.5);
      • inputting is_tax_unit_head for the 16-year-old gives that result.

      I ran that input: sc_tanf_dependent_child [T, F, T], gross earned [0, 500, 0], countable 250, sc_tanf 347.

   3. sc_tanf_countable_income.py:24-25: "Section 8.1 says "Do not round at any step of the TANF calculation", so countable income keeps its cents." CONFLICT_COMMENT_RE doesn't match it, because "8.1 says" names none of its source nouns. test_source_conflicts.py passes 103 with no KNOWN_UNMARKED change. None of this round's comments in variables/ matched either.

   4. Peak memory: see Tests.

7. Axiom. The draft is below. I didn't edit any issue, and rulespec-us is outside this session's GitHub scope, so I couldn't read NYC taxable income #1728.
8. Mutants. See Mutants.
9. Tests. See Tests and CI.
10. Body. I brought every section to this head: rules, expected values (new rows for Cases 22, 31–33, 36 and 37, and the gross limits), invariants, mutants, tests and overlap. The line is now impact: pending a comparison on the default dataset for 2026, main against this branch., and the axiom line is unchanged. I read the body back after the PATCH.

Tests at the final head

One process per file, never more than four at a time:

• YAML. The 15 SC TANF files pass, 145 cases: integration 37, gross_income_eligible 21, gross_unearned_income 18, sc_tanf 16, need_standard 12, dependent_child 9, and countable_earned_income_person, countable_income_eligible, eligible, income_eligible, resources_eligible and ssi_recipient 4 each. benefit_group_size and countable_resources have 3 each, and gross_earned_income 2. They ran at 675ca2d115, whose code and YAML match the head. This round changed integration, dependent_child, gross_income_eligible and income_eligible.
• Python. test_sc_tanf_child_earnings_support_invariants.py passes 4 at the head; it's the one file the last commit changed. test_sc_tanf_unearned_income_invariants.py passes 2, code_health/test_source_conflicts.py 103, test_reference_urls.py 32, test_no_concatenated_reference_urls.py 48 and test_known_dead_reference_urls.py 40.
• Commit 1 on its own (a1cc4925ae): integration 36, test_source_conflicts.py 103 and the property file 4 all pass.
• Peak memory (suggestion 4). GNU time wasn't installed, so I installed it (time 1.9, via apt). /usr/bin/time -v on the property file at the head gave 43.5 s wall clock including startup, 7.7 s of tests, and a peak RSS of 1,054,820 KB. At 96bd29a40f, before this round, it gave 43.5 s and 1,052,088 KB. Both are local runs on a 4 vCPU, 16 GB VM with Python 3.14. The file runs in the Makefile's remaining Python group, which the Makefile records at 14.4 GB and 22:18 in CI run 38141803955.
• Lint. ruff format --check and ruff check pass on the PR's nine Python files (ruff 0.15.5).
• Not run. I ran no microsimulation and no full suite, and nothing needed private data.

Mutants

Every row ran this round in a scratch worktree. Rows 1–12 are earlier rounds' mutants, rerun at the final head e4e6861e6a against integration.yaml, sc_tanf_dependent_child.yaml, sc_tanf.yaml, sc_tanf_need_standard.yaml, sc_tanf_gross_income_eligible.yaml, sc_tanf_countable_income_eligible.yaml, sc_tanf_gross_earned_income.yaml and the property file. Rows 2 and 4 are rewritten for the per-payer deduction. Rows 13–15 ran at 675ca2d115, whose code and YAML match the final head, against all 15 SC TANF YAML files and both property files. Rows 16–21 ran there against integration.yaml and the property file, plus both gross income files for rows 19–21. Row 20's property run and row 15's case names are from the final head. git status was clean after each.

#	Mutant	Result	Killed by
1	Drop the dependent-child exclusion	killed	integration.yaml Cases 19, 20, 30, 31; sc_tanf_dependent_child.yaml Cases 4, 8, 9; property (c)
2	Take each payer's support from that payer's gross earnings (floored at zero), apply the §8.12 disregard, then take the rest from unearned income	killed	integration.yaml Cases 21, 23, 25, 35, 36. Case 25 (January) is the ordering case; Case 21 fails on its asserted deduction, since this mutant moves the support into earned income.
3	Let child support reduce sc_tanf_gross_income	killed	integration.yaml Cases 21, 23, 25 (gross income); property (e)
4	Limit each payer's deduction to that payer's earned income	killed	integration.yaml Cases 23, 33
5	Drop the need standard floor	killed	sc_tanf_need_standard.yaml 11 of 12; integration.yaml Cases 1, 2, 5, 6, 15, 18, 26–29, 32; sc_tanf.yaml Case 8; sc_tanf_gross_income_eligible.yaml Case 5 and the eight "one cent above" cases; sc_tanf_countable_income_eligible.yaml Case 4
6	Apply the $10 test before flooring the benefit	survived (equivalent)	⌊x⌋ ≥ 10 exactly when x ≥ 10, so the order can't change a payment
7	Treat a pregnant adult as a dependent child	killed	integration.yaml Case 24; sc_tanf_dependent_child.yaml Case 3; both (f) differentials
8	Drop the no-adult clause (heads_household = head_or_spouse)	killed	sc_tanf_dependent_child.yaml Cases 5, 6, 7, 8; integration.yaml Cases 26, 28, 29, 30
9	Every member of a group with no adult heads it (oldest replaced with True)	killed	sc_tanf_dependent_child.yaml Cases 5, 6, 7, 8; integration.yaml Cases 26, 28, 29, 30. Case 2 no longer kills it, because an input head now turns the rule off.
10	Drop the head test (return meets_age_test)	killed	sc_tanf_dependent_child.yaml Cases 2, 5–9; integration.yaml Cases 26–31
11	Pick one oldest member by rank (oldest = person.get_rank(person.spm_unit, -age) == 0)	killed	sc_tanf_dependent_child.yaml Cases 6, 7; integration.yaml Cases 28, 29
12	Make every member within a year of the oldest a head (oldest = age >= person.spm_unit.max(age) - 1)	killed	sc_tanf_dependent_child.yaml Case 8; integration.yaml Case 30
13	Drop the no_head term (head_or_spouse | (no_adult & oldest)), the reading before this round	killed	sc_tanf_dependent_child.yaml Case 9; integration.yaml Case 31
14	Pool the deduction across the group: min_(sum(support), max_(sum(income), 0))	killed	integration.yaml Cases 22, 32, 36
15	Restore the floor on the gross income limit	killed	sc_tanf_gross_income_eligible.yaml Case 4 and the eight "one cent below" cases; integration.yaml Case 37
16	Restore the SSI payer condition	survived (equivalent)	An SSI recipient's counted income is already 0, so the cap gives 0 with or without it. Case 32 pins the result.
17	Restore the dependent-child payer condition	killed	integration.yaml Case 33
18	Drop the max_(child_support_expense, 0) clamp	killed	integration.yaml Case 34
19	Floor countable income at 0 instead of the members' negative incomes	killed	integration.yaml Cases 35, 36
20	Drop the countable income floor (income - deduction)	killed	property (b), on the fixed January example. At 675ca2d115, before that example, random draws missed it and it survived.
21	Restore the old where(deduction > 0, max_(income - deduction, 0), income)	killed	integration.yaml Case 36

The property tests killed rows 1, 3, 7 and 20 only. Every generated group has an adult head, only adults pay support, and no member has a loss. Row 13 was the body's mutant 13 before this round: the code now takes that reading, so the reverse mutant is the check.

The property tests found one real bug in round 1: countable income came out a float32 rounding error below zero when a deduction used up a payer's income. The stored float32 deduction can exceed the float64 income sum by about 3e-5. sc_tanf_countable_income floors the result at the sum of the members' negative incomes. That sum is zero unless a member has a self-employment or farm loss. Generated groups have no loss and no negative income, so the floor leaves (f) exact.

CI

Pull request run 42641205118 is the only run at e4e6861e6a. It concluded success at 16:14:08Z: 36 of 36 jobs passed, and none was skipped or cancelled. At 16:06Z, 60 minutes after the 15:06Z push, three Contrib shards were still running; they passed by 16:14Z. The run included Microsimulation and Household API Partners.

Judgment calls

1. I replaced the countable income branch, not just pinned it. Under a per-payer deduction, the old where(deduction > 0, max_(income - deduction, 0), income) misbehaves in one case: a member's self-employment loss in a 50%-disregard month outweighs what's left of a payer's income.
   • There it reports 0 instead of the negative total, so countable income rises as support rises.
   • sc_tanf_countable_income.py:31-33 now returns max_(income - deduction, sum of min(member income, 0)).
   • Each payer's deduction stops at that payer's income, so in exact arithmetic the total never falls below the members' own losses. The floor absorbs only the float32 residue, about 3e-5, between the stored deduction and the float64 income sum.
   • Case 36 gives −500 where the old branch gives 0. Both readers treat any total at or below zero alike (countable < need; the payment caps at need × 0.4046), so sc_tanf is 449 either way.
2. I removed both payer conditions (item 5). The SSI condition was equivalent. The dependent-child condition contradicted 114-1140(L)'s "a budget group member", which the C1 ruling follows; Case 33 now pins that reading. If the hub reads the C1 ruling as adults only, restoring * ~person("sc_tanf_dependent_child", period) on support_paid and flipping Case 33 to 368 undoes it.
3. The new candidates' years are from: 2026, federal-fiscal, with no to. Vol. 67 took effect August 1, 2026, and neither conflict depends on a guideline year. The above-eight entry keeps 2026 to 2026, because its $229 restates the 2025 guideline.
4. I added two tests the brief didn't name.
   • Case 37 gives the axiom draft's gross-limit companion a YAML twin.
   • The fixed property example makes row 20 deterministic: the no-floor mutant survived the random draws at 675ca2d115.
5. I rewrote rows 2 and 4 for the per-payer code and reran rows 1–12 at the head, so every "Killed by" cell in the body is from this round. Row 9 lost Case 2 as a killer, since an input head now turns the no-adult rule off.
6. I split C1, C2 and S1 into separate commits by staging intermediate versions of integration.yaml and candidates.yaml. Composing them gives the final files byte for byte.
7. I pushed once, at 15:06Z, before the rows 1–12 reruns finished. The code was final, and those reruns feed only the body; none needed a new case.
8. I fixed sc_tanf_income_eligible.yaml's comments though the round doesn't otherwise touch that file. They still described the unfloored need standard (851.67) and limit (1,575.58).
9. I left two things as I found them.
   • The body's dataset sentence ("all 29 people in its 28 SPM units with no adult hold the head role") is the hub's earlier measurement. I can't read the dataset.
   • Main's two pre-existing out-of-order pairs in candidates.yaml stay; this PR doesn't create them.

Draft refreshed extension for TheAxiomFoundation/rulespec-us#1728 (for the hub to post)

#10581 now follows S.C. Code Regs. Chapter 114 where it differs from TANF Policy Manual Vol. 67. This refreshes the 2026-11-02 extension, written at cdf464ad90, to head e4e6861e6a. Quotes are from the July 2025 Chapter 114 compilation and Manual Vol. 67 (sha256 cac6566b…), page numbers as PDF file pages.

Law to add to the corpus rows

• 114-1140(L) (Chapter 114, p.212): child support payments "made by a budget group member to legal dependents who are not budget group members may be deducted from the income of the payor."
• 114-1140(N)(1) (p.213): "The earned income of dependent children is excluded in the FI gross income limit test and the need and benefit determination."
• 114-1150(A) (p.217): gross income "must not exceed 185 percent of the need standard by family size."
• 114-1130(E) (p.204), optional: a dependent child "must be under eighteen years of age or under nineteen years of age and a full-time student in a secondary school".

114-1140(A)(4) stays as quoted. It rounds down only "the standard of need or the payment amount".

Readings that change

• Payer (replaces "Adult payer"). Support that any budget group member pays to legal dependents outside the group comes off that member's own income. It comes off earned income after the §8.12 disregards, then unearned income (the order of §8.16 C, D and F), never below zero. The deductions are summed over the group. A dependent child can be a payer; the child's earnings are excluded, so the support comes off unearned income. §7.7 excludes an SSI recipient's income, so an SSI recipient's support is never deducted. The manual's "any adult" pooling (§8.16 D and F, §8.11 "as income to the BG") gives way to the regulation.
• Gross income limit. The limit is 1.85 times the need standard with its cents. The §8.15 note's "drop the cents" gives way to 114-1150(A); only the need standard and the payment are rounded down.
• Dependent child's earnings. They are excluded from the gross income limit test as well as the net test and the benefit (114-1140(N)(1)).
• Head of household stays an input. PolicyEngine's oldest-member rule for groups with no adult applies only when no member holds a head or spouse role, so it is an inference for missing roles and needs no RuleSpec counterpart.

Companion tests (hand-computed from the law; each has a YAML twin in policyengine_us/tests/policy/baseline/gov/states/sc/tanf/):

• Test 7, now $246 (integration.yaml Case 22). May 2026. Adult A pays $2,400 a year and has no income; adult B earns $600 a month; one child. A has no income to deduct from, so the deduction is $0. Countable income is 600 − 100 = 500, and the payment is ⌊(1,110 − 500) × 0.4046 = 246.81⌋ = $246.
• New: dependent-child payer, $408 (integration.yaml Case 33). May 2026. Parent, a 17-year-old dependent child with $2,400 a year of survivor benefits who pays $1,200 a year of support, and a 10-year-old. The deduction is min(100, 200) = 100 off the child's unearned income, so countable income is 100. The payment is ⌊(1,110 − 100) × 0.4046 = 408.65⌋ = $408.
• New: SSI payer, $154 (integration.yaml Case 32). May 2026. An SSI-recipient parent pays $2,400 a year; the other parent earns $600 a month; one child. The SSI parent leaves the group, so its size is 2 and the need standard ⌊21,150 / 12 × 0.5 = 881.25⌋ = 881. The payer has no counted income, so the deduction is 0 and countable income is 500. The payment is ⌊(881 − 500) × 0.4046 = 154.15⌋ = $154.
• New: gross limit with cents, $26 (integration.yaml Case 37). January 2026. A parent with $19,554 a year of wages ($1,629.50 a month) and one child. The limit is 881 × 1.85 = 1,629.85, so the group passes. The 50% disregard leaves 814.75, and the payment is ⌊(881 − 814.75) × 0.4046 = 26.80⌋ = $26. With the cents dropped, the limit is 1,629 and the payment is $0.
• New: gross limit boundary (sc_tanf_gross_income_eligible.yaml). January 2026, group of two: gross income of $1,629.84 passes and $1,629.86 fails.
• New: input head of household, $449 (integration.yaml Case 31). January 2026. A 16-year-old head of household; a 17-year-old who is that head's tax dependent and earns $500 a month; a baby. The 17-year-old is a dependent child, so 114-1140(N)(1) excludes the earnings. Countable income is 0, and the payment is ⌊1,110 × 0.4046 = 449.11⌋ = $449.

Need standard reconciliation. The 2026-11-05 review reports that the issue body's need-standard output and its companion test 11 still give the unfloored $851.67. Under 114-1140(A)(4), the January 2025 need standard for two is ⌊(15,060 + 5,380) / 12 × 0.5 = 851.67⌋ = $851. That matches the extension's floored $851 and its $0 companion case for countable income of $826.80 (⌊(851 − 826.80) × 0.4046 = 9.79⌋ = 9, under the $10 minimum). Change test 11's expected value to $851, and drop the note that the standard keeps its cents.

Proposed follow-up issues

• SC TANF: stop a self-employment loss from offsetting other income in the 50% disregard months. sc_tanf_countable_earned_income_person floors each person at zero after the $100 disregard but not after the 50% one. So a January–April loss offsets other members' income and can take countable income below zero (Case 35); check §§7.1 and 8.16 C and treat both periods alike.
• SC TANF: move the remaining Volume 65 references to Volume 67. Four SC TANF variables (sc_tanf_gross_income, sc_tanf_countable_earned_income_person, sc_tanf_countable_income_eligible, sc_tanf_income_eligible) and nine parameters still cite Manual Volume 65. Repoint them to Volume 67, and to Chapter 114 where a regulation governs, such as 114-1150(A) for gross_income_limit.yaml.

---

Generated by Claude Code