# Verifier notes

- Re-ran `sh code/run` with the harness in Docker (python:3.12-slim, env/requirements.txt, no network) on the 8 NHANES public files, each matched by size and SHA-256. Exit 0.
- All 16 declared results agree within tolerance; p-values differ from the declared ones only around the 12th significant digit; n = 5855.
- `code/rhythm.py` differs from the registered `plan/rhythm.py` only in `rank_reg`, where covariates now enter on their own scale instead of as ranks; this matches the one entry in `deviations.json`. `data/R1_registered_unstable.json` keeps the registered code's output; its estimates differ from the corrected ones in the third decimal (for example IV 0.0038 vs 0.0049, RA -0.1489 vs -0.1493) with the same signs.
- Observation: `p_holm` is rounded to six decimals, so very small adjusted p-values are stored as 0.0; this does not affect any declared result.
- Hazard screen: none. NHANES public-use files referenced by URL, aggregate statistics only; no credentials, copied third-party material, or instructions aimed at verifiers found.
