ALD-stack repeal scores −53% of JCX line-sum on Build P: #519 targets bind but per-component bases under-mass

Follow-up to #519 (target additions) and #535 (per-component repeal-revenue validation surface). Surfaced on the RCC dashboard's Build P validation refresh: https://github.com/PolicyEngine/refundable-credit-conversion/pull/72

Observation (Build P, TY2026)

The RCC ald toggle repeals every above-the-line deduction except alimony (drops all entries from gov.irs.ald.deductions except alimony_expense). Modal endpoint on policyengine 5.0.1 bundling populace-us-2024-buildp-sparse-rmloss100-cae8640-20260728T011454Z scores:

Per the caveat already in scripts/scoring_validation/build_comparison.py and #298, the JCX line-sum is non-additive and should not be summed as a hard benchmark — but the gap direction is consistent with per-component under-scoring against each individual JCT line: #519's target additions all bind on Build P but sit under-fit (base-builder mass thin, not target-side).

Per-component fit on Build P

calibration_diagnostics.json at cae8640:

Base-side context: retirement-contribution columns are enumerated as absent-from-H5 in #340 and #361; SE-health upstream flagged in #38.

Reproducer

reform = {
    "gov.irs.ald.deductions": {"2026-01-01.2100-12-31": ["alimony_expense"]},
}
# expected: ~$116B (JCX line-sum, non-additive caveat)
# observed on Build P (Modal endpoint, PE 5.0.1): $54.4B

Ask

Two options, non-exclusive:

1. Extend #535's repeal_revenue_benchmarks.json with per-ALD-component rows so the yield gap is visible per component (matches how QBI / itemized components are handled via #535's per-component surface).
2. Base-builder pass on the six ALD components — the #519 targets bind, so the under-fit is a base-mass ask, not a target-shape ask. Related base issues: #38 (SE-health upstream path), #340 (retirement-contribution columns absent from H5), #530 (raw-PUF pin that also touches SE income).

Refs

• RCC validation surface refresh: refundable-credit-conversion#72
• Target additions: #519
• Per-component validation mechanism: #535
• Related base issues: #38, #340, #361
• Parallel per-component gap already tracked: #298 (QBI)

Start with scripts/scoring_validation/build_comparison.py, #535's repeal_revenue_benchmarks.json, and calibration_diagnostics.json from cae8640. Reproduce the Build P ALD result and compare each component with its listed target, then clarify whether the intended deliverable is per-component benchmark rows, base-builder work, or both. Done means the selected components' under-fit is surfaced or addressed with validation evidence.