Count Missouri TANF parents claimed as tax dependents as unit members

[MaxGhenis]

This change includes a biological or adoptive parent in Missouri's TANF assistance unit when someone else claims that parent as a tax dependent. In the target household—a marked non-parent grandparent, a dependent adult mother, and her child—the mother and child form the unit. The mother's income counts and can reduce or eliminate the children's grant. Max's d1049 ruling (2026-10-08) makes explicit parent_1_id / parent_2_id inputs the first source for parenthood, retaining the 12–50-year age-window imputation only for dependent children whose two parent ids are unknown (0).

The implementation retains main's #9741 get_override_branch helpers. Coverage checks income loss and its zero-income counterpart, the own-tax-unit child guard, the complete non-parent-caretaker identification vector and converse, annual input overrides, and the January snapshot. A separate-return biological-parent case also records the remaining cross-tax-unit membership limitation.

axiom: TheAxiomFoundation/rulespec-us#1444 queued (still open; this does not claim that the dependent-parent rule has been encoded).

Rule and retained interpretation

Parents of eligible children are mandatory members regardless of tax dependency. SSI recipients are excluded, and a biological or adoptive parent's presence excludes a non-parent caretaker. These rules follow 13 CSR 40-2.300 and 40-2.310 and Combined IM Policy Manual 4.2.

Max's d1049 ruling (2026-10-08) approves interpretation (i) for an 18-year-old parent enrolled in secondary school: the 18-year-old, her baby and her own parent form one combined unit, size three in the zero-income example. This is the approved modeling interpretation rather than a claim of settled law. Both archived 0210.005.30 and current 4.3 define minor parent as under 18, including the birthday month. Alternative readings give size two; separate grouping and major-parent deeming remain outside this build.

The default mo_tanf_is_parent_of_dependent_child first reads each relevant dependent child's parent_1_id and parent_2_id. A known id naming the person establishes parenthood regardless of their age gap; known ids naming other people do not establish parenthood. Only a child with both ids unknown (0) uses the existing 12–50-year age-window fallback, using the existing is_parent signal. The fallback imputes a relationship; it is not an eligibility rule. The same precedence applies to the presumed parenthood of unmarked tax-unit heads and spouses in membership and non-parent-caretaker identification: known ids identify named adults without distinguishing parent type, while unknown ids retain the existing proxy. This follows Max's 2026-10-07 principle to "make inputs as leaf-nodey as possible": callers can supply the underlying parent ids instead of overriding a derived parenthood flag. The flag is annual and uses January's dependent-child status: changing monthly_age later in the year does not recalculate that year's flag.

The d1049 implementation prefers explicit parent ids and preserves the existing unknown-id fallback, including the grandmother regression. A direct read-only check of the pinned Populace HDF at snapshot 9a814a3b3b53c0ecd6e1737b6ec862c31300ef6f verified SHA-256 6496cc4393d4d3c6574f76eca231de5898c803b9067645591fd5c4d3e65aee84: parent_1_id and parent_2_id are both absent from its 318 person columns and from every other entity table (166,321 people). Raw CPS PEPAR1 / PEPAR2 columns exist, but the loader does not convert them to these PolicyEngine inputs. Both leaf inputs have no formula and default to 0; dataset loading and uprating preserve the existing columns. Thus d1049's three Missouri formulas take their unchanged unknown-id path for this dataset. No new microsimulations were run.

An explicit flag overrides the default for the supplied year. Supplying it for one person makes unspecified people false for that year, so callers should supply it for every applicable parent. Other years can still use the formula. An explicit true does not add a real biological or adoptive parent whose own tax unit has no dependent child, even when another tax unit supplies the SPM unit's eligible child. The formulas group children within tax units, and this boolean input cannot attach a child from another tax unit. Mandatory membership for a separately filing parent under 13 CSR 40-2.300(5)(C) therefore remains unmodeled; the new YAML case explicitly verifies that the flag is true while that parent remains excluded.

Caller contract and remaining relationship limits

The shared parent_1_id / parent_2_id inputs name natural, adoptive and step-parents alike; they do not record parent type. Missouri's automatic parent flag consequently treats a named step-parent as biological or adoptive, incorrectly adding their needs and full income and potentially blocking an NPCR. Only natural or adoptive parents are mandatory and block an NPCR under 13 CSR 40-2.300(5)(C)/(D); 40-2.310(8)(B)1.D.(II) instead deems a step-parent's income after deductions. This round discloses that limitation, as permitted by review r3 P2; it does not implement relationship typing or step-parent income deeming.

Preserve the shared parent ids for other programs. When ids include a step-parent, explicitly supply mo_tanf_is_parent_of_dependent_child for every person for that year: true for every applicable biological or adoptive parent and false for step-parents and other people. Supplying one person's flag makes every unspecified person's flag false for that year. This corrects parent membership and NPCR identification when each relevant child's links are known. If any dependent child has both ids unknown, the head/spouse proxy remains active even with an explicit false; that inference limitation remains.

Known child parent ids do not identify a non-parent caretaker relative. Callers must explicitly mark an NPCR with mo_tanf_is_non_parent_caretaker; a grandmother with a grandchild naming an absent parent otherwise remains outside the child-only unit. Marking her permits consideration of her needs, subject to neediness, opt-out and other eligibility rules. This addresses review r3 P3.

Tests and hand-computed expectations

Review r3 adds three paired-year input-contract scenarios: a named step-parent's exclusion from membership and member income under a complete annual override; the same override restoring an NPCR when the natural parent is absent; and a known-absent-parent grandmother changing from child-only to consideration as NPCR when explicitly marked. The automatic named-step-parent outcomes remain asserted as evidence of the disclosed limitation. These fixtures verify the documented inputs, not step-parent deeming.

The new focused differential property restates the three b9e7fff948caf4f33313c942c09caa7c7ccc3dea formulas independently on 224 controlled households (784 people) per mode. With parent ids omitted or explicitly zero, invariants are unchanged age-and-own-child parent flags, unmarked caretaker presumption, parent-in-home NPCR exclusion, SSI exclusion, membership, unit size and zero-income payment. It covers age gaps 11/12/50/51, 18-year-old secondary-school status, zero/nonzero own-child counts, marked caretakers, spouses and SSI receipt. Existing property and differential coverage remains.

Review-documentation regressions are evidence-only audits: the reviewed head fails the P1 dataset/equivalence disclosure and P2/P3 caller-contract checks; the updated body and variable documentation pass. A documentation-only remedy cannot honestly create a failing-before/passing-after model result; the new behavioral fixtures characterize the unchanged model and its supported overrides.

D1049 adds YAML cases for an adoptive parent with a 55-year age gap and explicit parent ids, an unrelated adult 30 years older when the child's ids name someone else, and unchanged age-window behavior when both ids are 0. The property tests check that known parent-id relationships give the same parenthood result across age changes. The grandmother case and all existing cases remain.

R3 validation at 87d460fd1a6bf5fb8f7054fb7632f1234bd1ac15 on Python 3.13.9 / Core 3.32.15: 3 new paired-year YAML scenarios and all 29 expanded parent-property tests passed, including both 224-household unknown-id modes. Each file ran separately in the foreground. make format, lint and whitespace checks passed. The P1 and P2/P3 documentation audits failed against the preserved reviewed-head body/source and pass against this body and committed source. Formula ASTs are unchanged in this round, and the existing NPCR property file and all existing YAML cases remain intact. Partner files are untouched and partner tests were not rerun in this round. No microsimulations were run.

Saved validation at the reviewed d1049 head on Python 3.13.9 / Core 3.32.15: 13 parent YAML cases, 17 existing membership YAML cases, and 27 property tests passed. The new property spans 270 explicit-id combinations, including repeated ids in separate households. make format and lint passed. Each test file ran separately in the foreground; no microsimulations were run.

Against baf3fd187a16b0a77116f088a6d17ba1152d5027, the expanded parent YAML produced 10 assertion failures and 3 passes, and the new explicit-id property failed (1 failed, 26 deselected). Only main's two leaf input definitions were supplied as baseline prerequisites; the three Missouri formulas remained byte-for-byte unchanged. The named-head fixture was also rechecked with SSI explicitly zero and still failed on the baseline parent flag. The updated implementation passes all new cases. Original property functions and YAML cases, including the grandmother regression, are preserved; partner contract test expectations are untouched. Main was merged to reuse #9406's parent-id inputs and shared household-resolution helpers. Baseline failure receipts and final passing logs are retained outside the committed tree.

• The two-tax-unit guard case expects membership [true, true, false], size two, and $234.08628 (678 × 0.34526). It closes a coverage gap: main and the old PR head pass, while dropping has_dependent_child must fail.
• A marked grandparent, dependent mother, and three children with the mother's annual disability benefits of $14,400 produce size four and $1,200 monthly countable income. That exceeds the $990 need standard, so the grant is zero. Main excludes the mother and pays $292.08996 to the children.
• Changing only those disability benefits to zero preserves size four and pays $341.8074 (990 × 0.34526), $49.71744/month above main.
• The NPCR property independently asserts every person's identification flag and the existence/converse condition. Membership expectations still use the model's inclusion decision; they do not independently restate neediness or the grant comparison. Saved flag-on/flag-off property comparisons both ran the then-current code and did not execute pre-PR code.

R2 adds two annual-input YAML cases and a real separate-return parent case. The annual tests verify that an explicit 2026 override leaves the 2025 and 2027 defaults active, and that January's parent flag still governs July membership after the older child ages out. All people's monthly ages are supplied in the snapshot fixture, because a partial monthly-age input otherwise fills unspecified people with zero. The separate-return case expects parent flags [false, false, true], membership [true, true, false], size two, and $234.08628, documenting the unsupported legally mandatory father.

Historical r2 targeted validation ran one YAML file at a time on Python 3.14.4 and Core 3.32.8: the annual parent-input file passed both cases, and the membership file passed all 17 cases, including the real separately filing parent. Each of the three r2 cases passed. Formatting and lint passed at r2; its formula ASTs matched b9e7fff, and partner test files were untouched. These are historical r2 results, not checks of the d1049 formula.

Saved pre-r2 evidence passed 214 Missouri TANF YAML cases across 26 files, 66 property cases, four explicit-source-role checks, four input-definition cases, and 623 partner cases across 144 files. Both YAML folders ran with one worker, and partner expectations remain unchanged. The pre-r2 member file passes all 16 cases; removing has_dependent_child fails the intended guard case, while the original 14-case file passes that mutation. The saved properties detect both the guard removal and npcr_never; the original properties pass both mutations. Every intended failure is a saved call-phase AssertionError. The original PR head passes all 16 pre-r2 member cases; frozen main has five passes and eleven expected call-phase assertions, including the income-loss and zero-income pair.

Missouri non-parent caretaker (NPCR) behavior changes despite those unchanged partner expectations. Review r1 identifies three revised non-partner expectations, which it checked against 13 CSR 40-2.300(5)(C) and (D) and found to move in the legally correct direction:

• In the marked-grandparent/great-grandparent case, a great-grandparent aged 80 with own children in the home previously blocked any NPCR. Her 72-year gap to the child aged 8 fails the new parent-imputation window, so the marked grandparent is now included as NPCR: size two, $234.08628.
• The NPCR integration case changes the previously documented underpayment to size two and $234.08628.
• The NPCR property grid now expects adult_parent membership to be true rather than false. Its test is renamed to describe that current behavior.

The original evidence workflow remains failed: its case-sensitive detecting-name wrapper misses the authored guard title, and its two cross-checkout comparisons initially encountered pytest ImportPathMismatchError. The successful minimal continuation reran only those two comparisons from an exact temporary fixture outside the package checkouts. Combined saved-evidence validation is valid=true, errors=[], with completed controller provenance checked and 126 source files preserved; all other completed tests and simulations were reused byte-for-byte. The original failing receipts and wrapper exit remain preserved. See the compact artifact and full evidence report. Normal PR CI is a separate check.

Saved pre-d1049 performance evidence: the added Python coverage checks the complete NPCR identification vector and converse, while retaining the existing simulation-isolation grid. It belongs to the existing Rest/core CI group. On the sequential Linux runner, both property files took 55.25 seconds wall / 55.56 seconds CPU, peak RSS 1,242,112 kbytes. The old/new dependent-property commands, each including import/setup and three mutation sessions, took 86.08 / 85.80 seconds wall, with peaks 1,198,112 / 1,201,612 kbytes. Their unmutated pytest sessions were 24 cases in 14.55 seconds versus 26 in 14.57 seconds; per-session CPU/RSS were not separately measured. These are single-run measurements, not a measurement of the whole Rest group. Permanent CI routing and runner count are unchanged; all evidence runs are sequential.

Payment expectations retain the model's unrounded convention; payment rounding is a separate limitation.

Parent-pointer evidence

The precision and recall figures below record the saved age-window evaluation before d1049, not a new run of the explicit-id-first formula. The direct schema/digest check above now establishes that this pinned default dataset contains neither PolicyEngine parent-id input. For this dataset, d1049 therefore uses the same fallback relationship rule as b9e7fff; this is an equivalence inference by construction. The raw CPS pointers used for the historical evaluation are distinct from loaded parent_1_id / parent_2_id inputs.

The saved 2026-10-06 rescore uses the actual none school variant, rather than imputing school enrollment from A_HSCOL. A direct read of the pinned Populace HDF table's 318 person-column names confirms that A_HSCOL exists but is_in_secondary_school does not. The school input defaults to false in the model.

These figures measure agreement with resolved CPS parent pointers among the tax-dependent, non-child candidate population with a dependent child in its own tax unit. They are not a legal determination of parenthood. The Populace build is populace-us-2024-spm-20260909, snapshot 9a814a3b3b53c0ecd6e1737b6ec862c31300ef6f, SHA-256 6496cc4393d4d3c6574f76eca231de5898c803b9067645591fd5c4d3e65aee84.

Data and rule (none)	Selected records	Precision weighted / unweighted	Recall weighted / unweighted
Census ASEC 2023–2025 pooled: own children only	85	64.21% / 68.24%	100% / 100%
Census ASEC 2023–2025 pooled: age guard	71	74.82% / 78.87%	95.63% / 96.55%
Populace 2024: own children only	74	95.17% / 94.59%	100% / 100%
Populace 2024: age guard	71	99.999993% / 95.77%	99.999984% / 97.14%

In Populace, the guarded default retains three false positives and loses two pointer parents, all with very small weights. Missouri has only four selected records, three distinct source persons, and Kish effective sample size 2.92; pointer precision and recall are 100% in that thin sample. The pooled Census Missouri sample selects no candidates, so its precision and recall are undefined. The Census tax-unit artifact explanation is an inference from the observed structures.

The out-of-unit-parent statistic runs in this direction: 73% of parents in another tax unit from their child cohabit with the child's other parent. It does not establish that most cohabiting parents file separately. Parents whose own tax unit has no dependent child remain outside this implementation even when their child lives in the same SPM unit.

Unknown-id relationship regressions can raise or lower grants

The available child count and ages cannot establish the correct relationships when both parent ids are unknown. The retained fallback can wrongly add a dependent adult: with no income, their needs raise the grant; with income, counting it can reduce or eliminate the grant. These fallback regressions run in both directions. The table records saved pre-d1049 outcomes, which remain applicable to these fixtures with unknown parent ids; known parent ids now take precedence.

Household	Main	PR	Legal unit
Unmarked grandparent 55; dependent mother 20; child 3	Grandparent and child, size two, $234.08628	All three, size three, $292.08996	Mother and child, size two
Head 40 and child 10; dependent adult 45 whose own child is 20	Head and child, size two, $234.08628	Adds adult 45, size three, $292.08996	Head and child, size two
Mother/head 38; dependent grandmother 58 with $500/month UC; child 12	Mother and child, size two, $234.08628	Adds grandmother and UC, size three, $0	Mother and child, size two

The two zero-income cases each overpay $58.00368/month, or $696.04416/year. The UC case underpays $234.08628/month, or $2,809.03536/year: the grandmother's 46-year gap passes the 12–50 imputation window, and her income eliminates the grant. This is the PR's own grandmother-with-UC test. Callers can now supply the child's parent ids when those relationships are known; the derived parenthood flag can still be overridden explicitly. The grandmother case remains unchanged when the ids are unknown.

Impact and resources

The saved unmodified-dataset comparison reports $0 entitlement change and $0 takeup-adjusted/actual TANF change against frozen main. The direct dataset-coverage check establishes that d1049 adds $0 further change relative to b9e7fff for this exact pinned Populace dataset, both unmodified and under the saved NPCR/bank-asset interventions, which did not supply parent ids. This follows by construction from absent parent-id columns, zero-default leaf inputs, and the unchanged unknown-id branches; it is an equivalence inference, not a fresh microsimulation. All impact table figures remain historical paired measurements against the recorded frozen main, and no current-head comparison with today's main was measured. No new microsimulations were run; data containing supplied parent ids remains outside these population impact claims.

The excluded-grandparent resource defect predates this PR. A marked grandparent's separately owned $5,000 still counts against the otherwise asset-free mother/child unit, producing zero rather than $234.08628. The legal counterexample is preserved outside the passing suite with a prepared resource follow-up. Saved household checks also confirmed the opposite error on the then-tested main, original head and candidate: an included NPCR's spouse's $5,000 is dropped, leaving zero countable resources and a $234.086273 grant. Those resources must count under current Manual 10.11. Excluding every nonmember would be wrong because excluded financially responsible people can still contribute resources.

The completed source run compares frozen main 5b1d5bdf47044607341d76f3c932ba2e7b74a060 with candidate b9e7fff948caf4f33313c942c09caa7c7ccc3dea, through evidence checkout e39bb2c4c248ac4699032d7053f7429213014e75. Candidate and evidence checkout have identical model, test and changelog contents. All twelve simulations completed on Python 3.14.7 and Core 3.32.15, using the exact default dataset digest above. The full population was 166,321 people / 59,900 SPM units; saved Missouri diagnostics contain 2,892 people / 1,051 units. Each intervention applied to both codes, with identities, weights and intervention vectors checked before comparison.

Saved year/scenario	Annual entitlement change	Percent	January change	Actual annual MO TANF change	Entitlement records gaining / losing
2026 unmodified	$0	0%	$0	$0	0 / 0
2026 seven grandparents marked	−$20,456,426.63	−15.706373%	−$1,704,702.22	$0	0 / 1
2026 marked, bank assets zeroed	−$15,230,533.09	−9.463245%	−$1,269,211.09	$0	2 / 1
2025 unmodified	$0	0%	$0	$0	0 / 0
2025 seven grandparents marked	−$20,311,380.00	−15.559409%	−$1,692,615.00	$0	0 / 1
2025 marked, bank assets zeroed	−$15,122,540.56	−9.391461%	−$1,260,211.71	$0	2 / 1

The gain/loss counts are SPM-unit entitlement records from the saved paired main/b9e7fff comparisons, identical in every month in each scenario/year. Actual TANF records gaining/losing are 0 / 0 in all six rows. These saved impact runs were not rerun for r2.

Every month's payment, membership and failure partition equals January's within each scenario/year; the twelve monthly entitlement sums reconcile to annual totals. Actual annual Missouri TANF remains $5,497,666.85 in 2026 and $5,458,685.75 in 2025 on both codes; national actual TANF changes are also zero. The separately saved entitlement × takeup diagnostic is unchanged and is not substituted for actual TANF.

Four members are added in four units every month, with none removed: 22,417.537 weighted in 2026 and 22,258.586 in 2025. Under unmodified data, all four units fail income tests and three also fail resources. After marking, SPM 17211's mandatory mother's disability income ($1,290.69/month in 2026, $1,247.64 in 2025) eliminates the children's $292.089966/month grant; there is one new income failure and no new resource failure in each month. SPM 76134 still fails on its own parent's income after bank assets are zeroed. Zeroing bank assets lets SPM 1017211 gain $49.717438/month; SPM 1134143, with tiny weight, gains $98.399094. None of the four affected records takes up TANF. SPM 17211 and 1017211 account for 99.999232% of absolute annual change in the bank sensitivity. Complete drivers, failure partitions and receipts are in the evidence report.

For context, the saved 2026-09-29 study measured the following 2026 sensitivities under core 3.32.8 and branch 0771d2c498. These are historical results, not measurements against today's main. Its unmodified comparison used then-branch code with the parent flag off; its marked sensitivities separately executed pre-branch code 10e3f989ba.

Historical scenario	Annual entitlement change	January change	Takeup-adjusted change
Unmodified data	$0	$0	$0
Seven grandparents marked NPCR	−$20,456,424 (−15.7%)	−$1,704,702	$0
Same marks, grandparents' bank assets zeroed	−$15,230,525 (−9.5%)	−$1,269,211	$0

Four members, about 22,418 weighted, were added in the unmodified study. All four affected units failed the income tests, and three also failed the resource limit. Grandparents' income alone does not explain all the failures: the mandatory mother's disability income eliminates one children's grant after marking, while another unit retains income from its own parent. In the asset sensitivity another mother/children unit gains $49.71744/month. Two records weighted about 5,836 and 8,759 dominate the totals; none of the moved records takes up TANF.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (909176a) to head (87d460f).
⚠️ Report is 320 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff            @@
##              main     #9676   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            4         5    +1     
  Lines           76       106   +30     
  Branches         2         1    -1     
=========================================
+ Hits            76       106   +30     

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]

US + core hub merge audit, us#9676 at 87d460fd1a6bf5fb8f7054fb7632f1234bd1ac15 (squash): Missouri TANF assistance unit.

• Max's ruling d1049 (2026-10-08):
  • read parent_1_id / parent_2_id when set;
  • use the 12–50-year age window only when both are unknown;
  • interpretation (i) approved: an 18-year-old parent in school, her baby and her own parent form one unit.
• Reviews: delta review r4 APPROVE at this head, with no P1, P2 or P3. The r3 findings (impact verification, step-parent disclosure, NPCR marking) are resolved.
• Impact: zero change on the default dataset.
  • The r4 reviewer independently hashed and inspected the pinned HDF: neither parent-ID input appears in any of its seven tables, and the loader doesn't derive them. So every record takes the age-window fallback.
  • Unmodified entitlement and takeup impact was $0 before this round, and stays $0.
• Disclosed limitations: step-parent typing and deeming, with an annual override available; resource and cross-tax-unit limits.
• Gates checked live at merge:
  • gh pr checks exits 0;
  • the latest Pull request run at the head concluded success;
  • MERGEABLE;
  • not a draft;
  • no CHANGES_REQUESTED review;
  • the semantic-overlap guard;
  • head pinned.
• --admin: used only because a required approving review is the sole block.