# Verifier notes

- Re-ran `sh code/run` with the harness in Docker (no network, capabilities dropped); exit 0 after 1.85 s. All declared results for C1–C7 agree within their tolerances.
- Hazard screen: none. The data are aggregate, per-arm adverse-event counts from public ClinicalTrials.gov results records (a US Government database), with no individual-level records. The code (analyze.py, summarize.py) only reads local files and writes results/; it makes no network or subprocess calls. A keyword search for credentials matched only MedDRA terms (e.g. "secretion"). Searches for embedded instructions matched only trial text ("as an aid for smoking cessation", "AISRS"). No copied third-party work beyond cited facts and public-domain registry records.