# Verifier notes

`sh code/run` ran in rootless podman with no network, on an image built from `env/Dockerfile` (python:3.13-slim pinned by digest, numpy 2.5.4, scipy 1.18.1, pandas 3.1.0, pyarrow 26.0.0). Exit code 0 after 6 min 44 s, within the 45-minute limit. Not over budget. All 14 claims (C1–C14) reproduced; every declared result agrees exactly.

The first two harness starts exited 125: `--userns keep-id` tried to reuse a stale id-mapped overlay layer (`b86d72e6…`) left after a Podman reinstall, whose directory was gone. A direct `podman run` without `keep-id` worked. Removing the cached `sj-harness:e3713a01a171c824` image (id `ea620f444f35`) and rebuilding let the third run start and finish.

Hidden-content scan: nothing hidden in 127 text files (23 binary parquet/other files).

Hazard: none. The work is a pre-registered re-test of published protein Mendelian-randomization claims on public summary statistics and open datasets (Open Targets CC0, FinnGen and Pan-UK Biobank public summaries, 1000 Genomes European dosages released for open use). The bundle holds identifiers, codes, and offsets, not PubMed abstracts or Europe PMC full texts. No child-sexual-abuse material, no weapons or unauthorized-access capability, no private identifiable records or credentials.
