Base-builder imputation stages: CDCC adult-care leg + SE health ALD attribution (#451 items 1–2)

[MaxGhenis]

Base-builder imputation stages for the CDCC adult-care leg and the SE health ALD (#451 items 1–2)

Delivers the two Deliverable-2-class increments from #451 (input-coverage readiness quartet): both were structural zeros on every certified export — no adult/disabled-dependent care input existed at all, and the $339.6B bulk premium mass carried no self-employment attribution, so the §162(l) ALD computed exactly $0.

Item 1 — adult_care_inputs base-builder stage (CDCC adult-care leg)

New checkpointed stage adult_care_post_clone in tools/build_us_puf_support_base.py (after childcare_post_clone), producing two person inputs:

• is_incapable_of_self_care ← measured ASEC PEDISDRS == 1 ("difficulty dressing or bathing", the self-care difficulty item). This is the §21(b)(1)(B)/(C) qualifying-individual test and the §21(d)(2) deemed-earnings gate in PolicyEngine-US 1.764.6.
• pre_subsidy_care_expenses — the CDCC adult-care dollar leg (feeds care_expenses → cdcc_relevant_expenses).

Source decision (ASEC vs SIPP), with receipts

Read from the Census API variable dictionaries this session (ASEC 2024: 1,051 variables; SIPP 2023: 5,202 variables):

Concept	ASEC 2024	SIPP 2023
Incapable-of-self-care flag	PEDISDRS ("Disability - Difficulty dressing or bathing") — already ingested by the eligibility stage	ESELFCARE (identical item, corroborates)
In-household adult-care $	none — only child items (HCHCARE_VAL "Annual amount paid for child care", SPM_CHILDCAREXPNS)	none in scope — TDPCAREAMT is "money paid for the care of a former HH member", which fails §21's same-principal-abode requirement

Decision: ASEC, for both legs. The flag is directly measured. Because neither instrument measures in-household adult-care spending, the dollar leg is an explicitly documented same-instrument proxy: the measured ASEC work-related childcare expense distribution — the same §21 employment-related expense class, subject to the same §21(c) per-person cap — supplies (a) the paid-care usage rate (weighted share of positive-childcare SPM units among units with a child under the §21(d)(1)(A) age limit and positive earnings, measured ASEC channel) and (b) the positive level distribution (weighted quantiles). No dollar level or usage number is invented outside the measured donor distribution, and the proxy is declared in the stage manifest notes.

Statute-shaped support restriction

Expenses are assigned only where §21 can bind: units with a measured disabled qualifying individual (a dependent, or a married head/spouse — a taxpayer alone is never their own qualifying individual), passing the §21(d) earnings test computed exactly as the engine's min_head_spouse_earned binds: both spouses earning, or a §21(d)(2) floor-eligible spouse (incapable of self-care, or the measured full-time college student) deemed while the other spouse actually earns — deeming never rescues a unit whose only earner is the floor-eligible spouse. Selection is a seeded, weight-targeted permutation prefix over sorted unit ids (invariant to person-row order) at the measured usage rate; levels are a distribution-preserving weighted-quantile draw (zero-weight donors excluded, grid from the selected units' own cumulative weights) placed on the qualifying person with the smallest id. Fully deterministic given the build seed and invariant to frame row order.

Engine side-channels (enumerated, mechanism receipts read this session)

is_incapable_of_self_care is also read by SNAP/Medicaid work-requirement exemptions, TANF dependent-care deductions (MT/TX), CCAP (MN/LA), DC POWER, PR dependent test, and state CDCC analogs (MN/HI) — all move from all-False defaults toward the measured disability structure. Measured baseline deltas on the certified N frame are in the receipts below.

Item 2 — self-employed premium attribution (SE health ALD)

self_employed_health_insurance_premiums is formula-owned in the engine (an adds aggregation gated by defined_for = is_self_employed) and cannot ship as a column — the export adapter correctly refuses engine-computed names. The stage therefore ships the two pure inputs the engine intends, via a new deterministic final operation attribute_self_employed_health_premiums on the existing other_health_insurance_premiums release stage (it must run after the PUF premium QRF, so the placement adds no new call in tools/build_us_fiscal_refresh_release.py; that file carries exactly one minimal edit — see the register note below — and this branch will be rebased after PR #491 lands):

• health_insurance_premiums := the finished reported non-Part-B premium, exactly where combined Schedule C income (self_employment_income_before_lsr + sstb_self_employment_income_before_lsr) is strictly positive, the person is outside the Medicare proxy (age 65+ or any Social Security disability income), and the person is outside measured employer-sponsored coverage (has_esi — the conservative, incomplete proxy for the §162(l)(2)(B) subsidized-employer-plan month exclusion; eligibility through a spouse's or dependent's employer is not measured and remains documented residual overbreadth); zero elsewhere. Attribution never invents premium mass.
• is_self_employed := strictly-positive combined Schedule C income.

The engine then computes the §162(l) ALD as min(total_self_employment_income, premiums). The banked SOI Pub 1304 Table 1.4 TY2023 anchor (3,595,764 returns / $31.23B, ledger#105, in the v9.2 feed) binds through the existing measure maps once support exists — no calibration-loss machinery is touched (#492 doctrine: this is the support work).

Why the Medicare proxy guard

medical_expense_health_insurance_premiums uses the direct premium input when nonzero, else non_part_b + part_b·enrolled. Restricting attribution to people whose modeled Part B add-on is zero makes the direct input numerically equal to the decomposed value for every attributed person, so the statutory medical-expense concept (itemized medical, SNAP medical, SPM MOOP) is invariant — verified by an engine contract test. The excluded SE∩Medicare slice forgoes ALD mass conservatively; the calibration solve owns the level against the SOI fact.

Bare-input side-channels (enumerated from the engine graph): MO qualified-premium subtraction, MI household resources, OH unreimbursed medical deduction, CA Orange County GA, MO CCS income test — all move from premiums-invisible toward the measured premium structure for the SE slice.

Probes (the keogh pattern) and coverage manifest

Two new reform_coverage_probes (data, in the regenerated release_input_coverage_manifest.json), each neutralizing the shipped dollar leaf at 2024 law with budget_measure: income_tax, expected_sign: negative, floors set from the measured effects below. Four new required coverage leaves: pre_subsidy_care_expenses, is_incapable_of_self_care (red gate until the next base rebuild — the schedule_d precedent), health_insurance_premiums, is_self_employed (produced by any new release run).

Measured receipts (certified N frame c3e378a-20260722T010408Z, 166,321 persons / 57,240 households, release weights, build seed 0)

All numbers are from the post-round-2 rerun (the mechanisms changed: §162(l)(2)(B) employer-coverage guard; order-invariant selection). Both stages run end-to-end on the certified frame; both signal gates pass with every round-2 certificate (flag identity, statute structure, plausibility ceiling, attribution identity) at zero violations; both probes execute through the real reform-coverage smoke machinery.

SE health attribution (with the §162(l)(2)(B) has_esi guard):

• Attributed premium mass: $16.37B on 1,402 carrier rows / 3.57M weighted people — the weighted carrier count lands within 1% of the SOI 1.4 TY2023 return count (3,595,764). is_self_employed: 12,098 rows / 37.5M weighted (11.0% person share).
• Baseline self_employed_health_insurance_ald: $11.97B against the banked $31.23B SOI fact. Before the guard the raw ALD was $31.24B — a compositionally lucky match that included §162(l)(2)(B)-ineligible premium months (ESI-covered Schedule C people); the guarded surface is deliberately conservative, and the calibration solve owns closing the level against the banked fact over this support (no loss machinery touched, Loss shape: flat cap at 1.0 makes past-cap rows free dumping grounds (17 rows pushed out in Build N); bounded-tail experiment + past-cap census diagnostic #492).
• Probe effect (neutralize health_insurance_premiums, 2024 income_tax, baseline−reform): −$1.449B → floor set at $300M.

CDCC adult care:

• is_incapable_of_self_care: 2,067 measured PEDISDRS carriers / 5.998M weighted (1.76% person share — consistent with published self-care-difficulty prevalence).

• pre_subsidy_care_expenses: 137 carrier rows / 492k weighted, $3.41B donor-matched mass.

• Probe effect (neutralize pre_subsidy_care_expenses, 2024 income_tax, baseline−reform): −$153.9M → floor set at $30M.

• Weighted ALD-carrier tax units: 3,539,030 vs SOI's 3,595,764 returns (−1.6%) — both the premium-carrier and ALD-carrier counts land within ~2% of the SOI return count.

Baseline drift vs certified N (both stages on, measured, not estimated):

measure	certified	with stages	Δ
income_tax	$2,265.25B	$2,263.63B	−$1.61B
cdcc	$3.315B	$3.486B	+$171M
snap	$93.16B	$93.86B	+$697M
spm_unit_benefits	$1,800.02B	$1,800.82B	+$797M

The income-tax delta is dominated by the SE-ALD channel (−$1.449B); the CDCC lift combines the adult-care dollar leg with the §21(d)(2) deeming the measured flag enables on existing childcare expenses; the SNAP lift is the measured flag reaching the work-requirement exemptions the statute reads it in.

The income-tax drift is the SE-health ALD binding plus the CDCC growth; the SNAP/benefits drift is the measured disability flag reaching SNAP/Medicaid work-requirement exemptions (statute-correct on measured data; levels remain owned by the calibrated SNAP targets at the next solve).

Register consistency (the one edit to the #491-shared file)

The release builder's US_DOCUMENTED_ABSENT_INPUTS register carried is_incapable_of_self_care with the claim "No direct ASEC item" — refuted by the dictionary receipts (PEDISDRS is the direct item), and now contradictory with the required-coverage leaf (the register-consistency gate correctly refused the pair). The entry is removed and the register docstring records that the adult_care_inputs stage seeds the flag, mirroring the is_pregnant precedent. This is the only change to tools/build_us_fiscal_refresh_release.py; the branch will be rebased once PR #491 merges.

Cross-family review

Two sol rounds (read-only, judgment tier), all findings fixed:

Round 1: plan/export/audit wiring completed (US_DONORS/US_STAGE_NAMES, the L0/refit export contract, builder summary + staged required_signals, runtime required-inputs mirror); SE gate hardened (null/non-boolean flags, nonfinite identity sources); SIPP receipt vintage-scoped (the 2018 panel's Child & Dependent Care module collected ECREPAYANYON/TDEPNDNTEXP; the 2023 PUF — the pinned donor vintage — does not release them); state consumers enumerated in the shipped notes.

Round 2 (5 blocking, 7 should-fix, 1 nit — full report in the PR thread): the §21(d)(2) work test was directional-blind (a unit whose only earner is the incapacitated spouse wrongly passed — now computed exactly as the engine binds, with a regression test); student-spouse deeming added from the measured college indicator; seeded selection made invariant to person-row order (sorted unit ids, smallest-person-id placement, order-invariance test); the weighted-quantile pairing now excludes zero-weight donors and grids on the selected units' own cumulative weights; the donor channel column is required rather than defaulting to all-ASEC; the adult-care gate became a real heal certificate (strict flag validation, PEDISDRS identity when present, statute-structure checks — carrier must be a qualifying person, one carrier per unit — and a documented plausibility ceiling); the SE attribution gained the §162(l)(2)(B) employer-coverage guard (has_esi); string-typed flags now fail both gates (pandas-3 str dtype included); the builder failing-gate parametrization and invocation assertions now cover the adult-care stage; the §21(b)(1)(A) citation corrected. All measured receipts below are from a post-fix rerun.

Tests

• test_us_adult_care.py (new): manifest pins incl. the source-decision receipts, statute-structure binding, per-seed determinism, taxpayer-not-own-QI, fail-closed on missing sources and degenerate usage rate, gate rejections, heal path, live 1.764.6 CDCC chain (deeming, neutralization) contract.
• test_us_other_health_insurance.py (extended): deterministic attribution identity with both guards, Schedule-C-loss handling, fail-closed sources, gate identity/structural-zero rejections, legacy-surface rebuild-or-refuse, live 1.764.6 §162(l) chain incl. the defined_for gate and the medical-expense invariance receipt, shipped-probe pins.
• Base-builder locked stage order updated (adult_care_post_clone).

Fixes nothing silently: a release whose export drops either channel now fails the coverage gate and the reform smoke.

Refs #451 (items 1–2).

🤖 Generated with Claude Code

[MaxGhenis]

Cross-family review, round 2 (sol, read-only, judgment tier): 5 blocking, 7 should-fix, 1 nit on the round-1 state. Every finding verified against the engine/pinned sources before acting; all are fixed on this branch, with the measured receipts rerun afterward. Disposition:

1. §21(d)(2) deeming direction (blocking) — confirmed against min_head_spouse_earned: a married unit whose only earner is the incapacitated spouse wrongly passed my work test (the engine deems the eligible spouse and keeps the other spouse's actual zero). The work test now computes exactly the engine's binding condition — both spouses earn, or a floor-eligible spouse is deemed while the other spouse earns — with a directional regression test (test_deeming_never_rescues_a_unit_whose_only_earner_is_the_disabled_spouse).
2. Student-spouse deeming (should-fix) — cdcc_income_floor_eligible also covers full-time students, and the measured is_full_time_college_student reaches the engine's is_full_time_student aggregation (verified). Now a pinned stage source; test added.
3. Donor-universe honesty (should-fix) — the shipped notes and docstring now state precisely what is measured: SPM-grain units with any member under the §21(b)(1)(A) age and any positive member earnings, an approximation at donor grain, not the tax-unit minimum-earnings test.
4. Row-order dependence (should-fix) — selection now permutes sorted unit ids and places on the smallest qualifying person_id; test_expense_assignment_is_invariant_to_person_row_order shuffles the frame and asserts identical per-person output.
5. Weighted-quantile pairing (should-fix) — zero-weight donors are excluded from the level distribution, and the grid is the selected units' own cumulative-weight midpoints, making the recipient-weighted draw reproduce the donor-weighted distribution.
6. Channel provenance (should-fix) — person_support_channel is now a required, pinned stage source; its absence fails closed instead of silently labeling every row ASEC.
7. Adult gate as heal certificate (blocking) — the gate now validates the flag strictly (bool/numeric-{0,1} dtypes only — both reproduced breaks fail), rechecks the PEDISDRS identity when the raw column is present, certifies statute structure (every carrier a §21 qualifying person of its unit; at most one carrier per unit; explicit failure when the role/unit columns are absent), and applies a documented $250k plausibility ceiling (the 1e300 reproduction fails).
8. §162(l) overbreadth (blocking) — the attribution now excludes measured employer-sponsored coverage (has_esi), the conservative proxy for the §162(l)(2)(B) subsidized-plan month exclusion (spouse's/dependent's-employer eligibility is not measured; documented residual overbreadth). This is the largest substantive change: attributed mass drops from $48.09B to $16.37B on 1,402 carriers (3.57M weighted — within 1% of the SOI return count), baseline ALD from $31.24B to $11.97B vs the $31.23B banked fact. The round-1 near-exact ALD match was compositionally lucky and included statutorily-excluded months; the guarded surface is conservative and the solve owns the level. Probe remeasured at −$1.449B, floor $300M.
9. String flags (should-fix) — both gates now reject non-bool/non-numeric flag dtypes outright (including pandas-3 str columns, where "0" would parse truthy in the engine); tests cover NaN, string, and nonfinite-identity-source cases.
10. L0 fixture regression (blocking) — already fixed in bb56a6e (the review snapshot predated it); 48/48 pass.
11. Builder test lock (should-fix) — adult_care joined the failing-gate parametrization with frame-flow assertions on both sides of the stage, and the taxpayer-alone test now asserts the surviving unit actually receives the expense.
12. Locked-file scope (blocking-if-unwaived) — the lane instruction explicitly allowed touching build_us_fiscal_refresh_release.py with a rebase-after-Share one critical-target register between builder gates and the publish contract (no loss changes) #491 obligation; Share one critical-target register between builder gates and the publish contract (no loss changes) #491 merged and this branch is rebased on top. The one edit (removing is_incapable_of_self_care from US_DOCUMENTED_ABSENT_INPUTS, whose "No direct ASEC item" claim the PEDISDRS receipts refute) was forced by the register-consistency gate — required-and-documented-absent is a contradiction the gate correctly refuses.
13. §21(b)(1)(A) citation (nit) — fixed; the comment now notes the engine parameter's own 21(d)(1)(A) citation carries the same value.

CDCC adult-care numbers moved with the order-invariant selection: 137 carriers / $3.41B / probe −$153.9M (floor $30M). The PR body and shipped probe reasons carry the post-fix numbers; the earlier #451 receipts comment is corrected in a follow-up there.