Materialize Medicaid and CHIP take-up via anchored count-calibration (ACA stage pattern)

[daphnehanse11]

Context

The #315 take-up contract marks takes_up_medicaid_if_eligible and takes_up_chip_if_eligible as rate_unsourced: us-data's state rates (0.53–0.99) are model-relative ratios (MACPAC enrollment over modeled eligibility) that fail the provenance rule, and no administrative participation rate clears the bar. A fixed-rate seed stage (the #294 SNAP pattern) is therefore doctrine-blocked for these programs.

Meanwhile the all-True defaults are actively harmful:

• Add Medicaid and CHIP eligibility-to-enrollment diagnostics #170: the published 2024 release shows enrollment mechanically equal to eligibility (Medicaid 72.30M == 72.30M; CHIP 4.98M == 4.98M).
• Weight calibration targets medicaid_enrolled against CMS counts (fiscal_targets.py cms_medicaid rows), so with all-True flags the solver shrinks the eligible population toward enrollment counts — biasing weights for everything correlated with Medicaid eligibility.
• Assign data-construction inputs added by recent policyengine-us merges (SNAP ABAWD discretionary exemptions, CHIP take-up) #323 (item 2): policyengine-us#8679 gates household CHIP value on chip_enrolled = is_chip_eligible & takes_up_chip_if_eligible; the engine's gov.hhs.chip.takeup_rate (0.93) is wired to nothing and the deterministic-country-package convention expects populace to assign the flag.

The doctrine-compliant path already exists in-repo: the native ACA stage (aca_marketplace_inputs in source_stages.json) anchors on reported coverage and fills via assign_binary_from_rate + calibrate_binary_assignment to CMS count facts — no participation rate cited. Those ops are generic in source_runtime.py (parameterized output/anchor/eligibility/domain/weight), just cosmetically labeled ACA. This is also how Urban's HIPSM works: survey-reported coverage anchors, and take-up fills the gap to a CMS point-in-time enrollment snapshot.

Goal

A person-grain source stage materializing both flags: reported-coverage anchored, count-calibrated to the cms_medicaid state monthly-snapshot facts, with release diagnostics and a degeneracy gate — delivering the follow-up #170's acceptance criteria anticipate.

Suggested implementation

• Builder frame prep: engine simulation computes person-level is_medicaid_eligible / is_chip_eligible, mirroring the ACA _calculate_array pattern in build_us_fiscal_refresh_release.py (~line 1785).
• Anchor: has_medicaid_health_coverage_at_interview (already carried in cps_carried.py). Reported recipients who are model-eligible always take up; document the reported-but-model-ineligible decision explicitly (ACA masks the anchor by eligibility; SNAP anchors unconditionally).
• Fill: assign_binary_from_rate with an in-build state rate (CMS snapshot over weighted modeled eligibles, mirroring the ACA rate derivation) on stable person draws, then calibrate_binary_assignment to state enrollment facts with preserve_true_anchors, domain: is_medicaid_eligible, person weights. Generalize the ACA-specific labels/salts of the binary ops as needed.
• CHIP: blocked in M-CHIP states until CHIP enrollment targets are concept-mismatched: CMS total_chip_enrollment includes M-CHIP; model materializes separate CHIP only (20 states zero-support) #321's concept mapping lands — either calibrate CHIP fill only where separate-CHIP concept support exists, or sequence behind the M-CHIP/separate-CHIP ledger split.
• Contract: move Medicaid/CHIP from rate_unsourced to a new treatment value (e.g. count_calibrated: anchored assignment calibrated to administrative counts, no participation rate cited). This is a doctrine amendment to the Seed TANF and EITC take-up from administrative rates; engine-asserted take-up contract (#312) #315 enum and needs contract-owner review.
• Gate + diagnostics (per Add Medicaid and CHIP eligibility-to-enrollment diagnostics #170): eligible count, enrolled count, ratio, nationally and by state. Fail on the all-True landmine; distinguish legitimate saturation where modeled eligibles < CMS target (combined eligibles 77.28M < CMS 79.15M means some states saturate by eligibility undercount, not by flag degeneracy).

Acceptance criteria

• Both flags materialized and non-degenerate nationally and per state, except documented saturated states.
• Anchored reporters preserved through calibration; stable identity across repeated builds.
• Contract records the new treatment with doctrine text; take-up contract test passes against the pinned engine.
• Add Medicaid and CHIP eligibility-to-enrollment diagnostics #170 diagnostics emitted; the gate distinguishes landmine from saturation.
• Weight calibration on medicaid_enrolled runs against take-up-materialized enrollment rather than raw eligibility.

Enrollment-concept semantics (point-in-time vs annual) are declared in the companion doctrine issue.