feat(verification): distributional V1 graders + rolling baseline (clinical eval)

[jim4226]

Summary

Adds a second flavor of V1 grader for domains where the "good" criterion is a scalar metric over a held-out sample (Dice score, anatomical landmark RMSE, Hausdorff95, boundary F-score) rather than a binary pass/fail check — and links it to the loop via a persisted rolling baseline so the threshold tightens with each promoted artifact.

Prompted by a follow-up exchange with Michael Cohen on graders for outcomes-based eval in clinical workflows. Draft response in brain/responses/01-michael-reply-distributional-graders.md.

What's new

• csis/verification/distributional.py — three types (DistributionalSample, RollingBaseline, DistributionalThreshold) plus a distributional_grader(name=..., threshold=..., baseline=...) factory that pins into the same GraderRegistry the existing V1 graders use. F6 (pinned-source-hash drift), cross-checkpoint cert signing, V2 critic, and auditor why-doc all keep working unmodified.
• make_clinical_imaging_registry() — four-metric V1 set tuned for the 2D-X-ray-to-3D-bone-reconstruction problem at sub-mm precision (dice_score, boundary_f1, landmark_rmse, hausdorff_95), each with (floor, op, summary_stat, min_samples, max_regression, regression_stat).
• update_baselines_after_promotion() — the explicit link from this iteration's verified gain to the next iteration's threshold. Daemon calls it after coord.run_iteration() returns outcome == "promoted"; the next iteration's grader reads the updated JSON cold.
• brain/research/02-distributional-graders.md — writeup mapping the pattern to Open Research Question Add a local-backend adapter (vLLM, llama.cpp, or Ollama) #1 ("Verifier-grade for fuzzy domains") and V5 calibration.

The pass criterion that matters

A grader passes iff:

1. summary_stat of the sample (mean, median, p10, p90 — configurable) clears floor under op.
2. AND min_samples is met (defends against zero-sample fake-passes).
3. AND (if baseline + max_regression are set) the sample's regression_stat is not worse than the median of recent per-iteration regression_stats by more than max_regression. Median-of-tail-watermarks is robust to a single outlier promoted iteration.

The closure does NOT mutate the baseline at evaluate time — promoting the baseline pre-decision would let a failing artifact contaminate the very watermark its successor will be measured against.

Test plan

• 20 new tests in tests/test_distributional_graders.py
• Summary statistics correctness (mean, median, p10, p90, single-value edge case)
• Rolling-baseline window enforcement + JSON persistence (write → fresh process → cold-load)
• Median-of-p10s robust to a single outlier promoted iteration
• Higher-is-better and lower-is-better threshold paths
• Silent-regression case (mean still clears floor, p10 collapsed → regression check fails)
• "No baseline yet" first-iteration case
• F6 pinning + drift detection holds with distributional graders in the registry
• Clinical-imaging registry happy path + sub-mm-precision-broken failure case
• End-to-end: Coordinator runs the full 8-step loop with the clinical registry, the cert carries every distributional grader's full metrics dict, the artifact promotes, the baseline updates, and the next iteration's grader reads the tighter watermark.
• Full suite: 237 passing (217 → 237)

https://claude.ai/code/session_01XesuxhRkBxEyHGGY4GzqdX

---

Generated by Claude Code