[MaxGhenis]

The procedural-interventions increment (Sarah Esty's frame): what happens to the measured rate and the cost-share bill if counted error dollars fall by X in a targeted group — an accounting construction throughout, never a causal claim, labeled as such in the artifact, the memo, and every docstring.

What's here

Capture curve (analysis/model_capture.py): out-of-sample FY2024 capture by the committed error model — out-of-sample by construction (the script verifies FY2024 is the test year and absent from training before fitting) with the recomputed weighted ROC-AUC matching the locked 0.7666262908275983 to full precision. Oracle/model/random shares of counted error dollars at 1/2/5/10/20% of weighted caseload, nationally and per state. Nine sparse low-budget state cells where model < random-expectation are locked as observed, not smoothed.

Intervention grid (analysis/interventions.py): {oracle, model, self-employment, random} × {1, 5, 10}% weighted coverage (exact via one fractional boundary case) × {25, 50}% effectiveness, per jurisdiction: single-measurement tier probabilities and expected cost share through the existing simulator (new optional targeted_memberships pass-through; element-lever path untouched), plus the sustained FY2028–30 construction through the persistence exposure machinery (#83) — the single-measurement delta subtracted from every anchored horizon year, named as an assumption.

Dollars: official FY2024 state issuance joined (FNS annual state summary, sha256-recorded, totals reconciled to the dollar; $93.5B national). Example, Colorado at 5% coverage / 50% effectiveness, sustained FY2028–30 expected cost share: random $436.8M, self-employment $422.3M, model $367.7M, oracle $224.9M — the model captures about a third of the oracle-minus-random gap.

Disclosures: the OBBBA delayed-implementation provision is not modeled (near-term exposure above 13.33% overstated — flagged after reading Schanzenbach's deck, which bins it as temporarily 0%); the FY2024-official vs fixed-real threshold convention mix; the issuance source's unverified disaster-supplement question.

Verification

• 22 tests green across interventions/capture/persistence, including both raw regeneration value-locks and the new dollars-equal-share-times-hash-guarded-issuance lock in every cell.
• Implementation lane: sol (gpt-5.6-sol), spec-driven; integration, dollar layer, and disclosure pass reviewed and written by the session main loop. Additive-only changes to simulate.py/persistence.py (optional params; persistence artifact byte-identical, its locks re-run green).

🤖 Generated with Claude Code