Correct scenario_031's Medicaid annotations to the engine's California disregard

[MaxGhenis]

Summary

The case note, reference explanation and all 43 row annotations for scenario_031 head_medicaid_eligible (California, single head age 67, reference: eligible) explained the reference by subtracting a Medicare Part B premium from countable income. policyengine-us 2.15.17, the reference engine, subtracts no premium. This PR rewrites that text to the engine's actual mechanism.

It is wording only:

• no reference, exclusion, failure class or score changes;
• the frozen run directory is untouched.

The finding came from the Part B audit (#193). Details are in reference_audit/2026-10-05-medicaid-031-annotations/README.md.

What the engine does (read in 2.15.17 source this session)

• Category. SENIOR_OR_DISABLED. is_ssi_aged is age ≥ 65, so age alone meets the category condition, with no SSI receipt or disability needed.
• Adult expansion group. is_adult_for_medicaid_nfc requires age 19 to 64 and no Medicare eligibility, so this head is outside it. It is the only formula in variables/gov/hhs/medicaid that bars on Medicare.
• Countable income.
  • medicaid_optional_senior_or_disabled_countable_income counts SSI unearned income: $9,445.47 Social Security (in full), $13,608 pension and $800 IRA, for $23,853.47. The $1,165 of alimony paid is not subtracted.
  • The state disregard senior_or_disabled.income.disregard.individual['CA'] is $230 a month. It is passed to _apply_ssi_exclusions as general_exclusion, so it replaces SSI's $20 exclusion: $23,853.47 − $2,760 = $21,093.47.
• Limit. 1.38 × the 2026 guideline for one person ($15,960) = $22,024.80. Income at or below the limit passes.
• Assets. $4,200 of countable resources, under CA's $130,000 limit.
• Part B. The engine models a $2,434.80 Part B premium, but countable income does not read it. With the premium at 0, or Medicare enrollment off, countable income is still $21,093.47 and the result is the same.

The old text compared "about $21,400" with "about $21,600", which is 138% of the 2025 guideline. Gross income is $1,828.67 over the real limit, and gross less only SSI's $20 exclusion is still over it. So the models that skipped California's disregard still erred.

Root cause. The judge prompt behind the published text stated the wrong mechanism twice:

• the case's grounding line (untracked grounding.csv, pinned by GROUNDING_SHA256): "a non-MAGI pathway whose income counting deducts health insurance premiums (including Medicare Part B)";
• the old reference explanation.

No other case's grounding carries that phrase. This PR fixes the explanation. The grounding line should be regenerated before the next audit stage that re-judges this case.

Changes

• rewrites.json: 45 whole-field rewrites (1 case note, 1 reference explanation, 43 rows).
• scripts/apply_rewrites.py:
  • applies the rewrites and re-pins the three files' sha256 in paper/snapshot/20260501/manifest.json;
  • refuses a missing or ambiguous row, or a field that holds neither text;
  • round-trips the CSVs byte for byte;
  • with --check, writes nothing and confirms every rewrite is in force.
• scripts/engine_values.py → verification/engine_values.json: recomputes every figure on 2.15.17 + latest_final and asserts each relation the text states, under the modeled premium, no_part_b and not_enrolled.
• scripts/payload_diff.py → verification/payload_diff.json: rebuilds the payload from the frozen run with export_country, once with main's annotations and once with these.
• verification/row_review.json: per-row review record (quoted model evidence, problems found, edits after review).
• tests/test_annotation_rewrites.py: keeps the ledger in force, so a release rebuilding these files from an older base can't silently restore the old text.

Invariants (stated and checked)

• Wording only. Against origin/main, exactly 45 fields differ, all on scenario_031 head_medicaid_eligible. Keys, order, quoting, failure_source, failure_subtype and reference_suspect are identical (checked by a field-by-field diff; test_rewritten_rows_keep_their_classes).
• No score change. The two payload rebuilds differ in 132 paths, all in that one cell's text fields: 43 annotation, 43 caseAnnotation and 46 referenceExplanation. No modelStats, programStats, heatmap or other cell value moves.
• Base reproduction. The base rebuild reproduces the frozen data.json.gz, except claude-fable-5's four usage fields, which the drivers carry (CARRIED_USAGE).
• Mechanism. engine_values.py asserts each of these:
  • countable income = gross − 12 × $230;
  • the limit is 1.38 × $15,960;
  • countable income ≤ the limit < gross − $240;
  • the premium-off and not-enrolled readings give identical countable income and eligibility;
  • the published reference is 1.
• Subtypes kept. The decisive-error tally (29 income counting / 4 threshold / 3 asset / 7 category) equals the recorded subtype counts (29 / 4 / 3 / 7).

Not changed here, and why

• The published dashboard keeps the old text until a data release re-exports. The payload is built from the frozen run dir, which this PR may not touch. payload_diff.json shows exactly what a re-export will change.
• scenario_114 (Audit the state income tax withheld in federal SALT; propose three exclusions #191, Audit the Medicare Part B premium in medical expenses; propose excluding scenario_114's Virginia income tax #193): its two outputs are proposed for exclusion, so the engine check of their notes and rows is posted on Audit the Medicare Part B premium in medical expenses; propose excluding scenario_114's Virginia income tax #193 for the adopting release instead of rewritten here.
• Over-the-counter expenses in other notes (scenarios 016, 075, 121): split into a separate task.
• Tool value vs explanation trailer. On this output, claude-opus-4.8 and claude-sonnet-4.6 submitted 1 in the tool call, but their explanations end "value = 0" and the frozen prediction is 0. Current parsing treats the tool value as canonical. That is a scoring question, split into a separate task.

Coordination with release tooling

• Harden the GPT-6.1 Sol driver: pin kept sidecars, reference explanations, re-opened flags and the export receipt #190 adds test_the_committed_reference_explanations_are_20260929s, which pins the committed us_case_reference_explanations.csv at HEAD to release 20260929's bytes. Whichever of Harden the GPT-6.1 Sol driver: pin kept sidecars, reference explanations, re-opened flags and the export receipt #190 and this PR merges second will see that test fail until it is scoped. Suggested scoping: check the blob at the release commit, or allow rewrites listed in this ledger. Commented on Harden the GPT-6.1 Sol driver: pin kept sidecars, reference explanations, re-opened flags and the export receipt #190.
• The 20260930b re-judge driver (branch judge-isolation-20260930b, no PR yet) builds from release 20260930's committed annotation bytes. Its freeze refuses a working tree that differs from them, so it fails loudly rather than reverting this. It will need to treat this ledger as part of its base. The owning session has been told.

Tests

• pytest -m "not slow": see the comment below for the run on this head; the baseline on origin/main was 1563 passed, 8 skipped.
• ruff check . and ruff format --check . are clean.
• apply_rewrites.py --check: 45 rewrites in force.

Review

An independent Opus 5.5 Subfleet review is running against 19d28a54. Its report will be added as verification/independent_review.md.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
policybench-site	Ready	Preview	Oct 5, 2026 6:40pm UTC

[MaxGhenis]

Local run on 19d28a54 (Python 3.12, uv sync --locked --extra dev, the CI command pytest -m "not slow"): 1568 passed, 8 skipped, 14 deselected. The baseline on origin/main 8b4c0ca1 was 1563 passed, 8 skipped; the difference is the 5 new tests in tests/test_annotation_rewrites.py. ruff check . and ruff format --check . are clean. apply_rewrites.py --check: 45 rewrites in force. engine_values.py (2.15.17 venv) and payload_diff.py ran with every assertion passing.

[MaxGhenis]

Shipped in release dashboard-data-20261006 (PR #202, squash-merged as 9ce4ade8, now the latest data release and live on policybench.org). This PR's commits were folded into that release as a merge commit, so its content is on main; closing it rather than merging it a second time. Its scenario_031 Medicaid annotation rewrites ship with the release's re-export.