A frozen denial-overturn predictor, scored in public, deterministically, month after month.
This repository is a pre-registration. Before we knew a single outcome, we froze a predictor and committed the sha256 hashes of the documents that define it. Each month we score it against newly published, real adjudicated outcomes — and anyone can check that we did not move the goalposts after the fact.
A predictor of whether a health-plan denial, appealed to independent medical review, is overturned
or upheld. Config v1-backbone-only, git tag foundry-predictor-freeze-v1. It is pure
deterministic code with no model calls in the scoring path — two scoring runs are byte-identical, so
a score can never depend on which day it was computed or which model was rented.
The predictor reads pre-decision structured fields only. The fields written after a decision — the determination itself, review timing, and the reviewer's findings text — are never inputs to a scored prediction. That guard is structural (the scoring code errors if a post-decision column appears), not a promise.
California DMHC Independent Medical Review (IMR) determinations, obtained via the California Health & Human Services open-data portal — a public record of denial → appeal → outcome cases.
Read every number with this caveat (verbatim): IMR cases = denials appealed to exhaustion, a selected sample. Calibration there ≠ pre-submission accuracy. Every report carries this block verbatim. We do not oversell — internally or externally.
- Freeze. The model, the pass/fail bars, and the scoring rules are fixed and hashed before any
evaluation-period outcome exists. Those three documents are hashed in
MANIFEST.md. - Predict, then hash, then wait. Each month the frozen predictor emits its probabilities for cases whose outcomes are not yet public. That predictions file is hashed and the hash is committed here — the predictions themselves stay private.
- Score when outcomes publish. The private predictions are revealed and scored by deterministic join+math code. Anyone can re-hash the revealed file and check it against the hash committed earlier. A mismatch would be provable — which is the entire point.
- Judge against a floor recomputed from the same month. Each month's baselines are recomputed on that month's cases, so the predictor is measured against a floor drawn from the same sample it is scored on, never a stale number.
Two consecutive losing months → publication pauses, retool before month three. A month's "win/loss" verdict is decided against the pre-registered baseline margins, not by any single report.
Committed before any month has been scored, so it can never be relaxed after a bad month.
- First public evaluation: the frozen predictor is scored on IMR cases published 2026-07-08 → 2026-07-31 (Gate 0), against the pre-registered margins.
- Numbers stay private for now. DMHC/OPA commercial-use approval is pending. Until it clears, only the pre-registration hashes are published here — no IMR-derived scores or figures. When the posture clears and each month is scored, the results and revealed predictions land in this repo.
| File | What |
|---|---|
MANIFEST.md |
sha256 of the three frozen pre-registration documents + a manifest hash over the ordered list. Hashes only — the documents' contents stay private until the OPA posture clears. |
Methodology, model card, and Gate-0 pre-registration are held privately at freeze and hashed above; they publish when the legal posture clears. This README is a summary; the manifest hash is the commitment.