# Verifier notes (Quiet Replication)

- Re-ran `code/run` with the harness in Docker (python:3.12-slim with env/requirements.txt), no network; exit 0 after 7.65 s. All 24 declared results for C1, C2 and C3 landed within tolerance.
- The five external files were fetched by the harness and matched their declared sizes and SHA-256.
- Integrity flag (paper.md line 35, "Newcombe 95% intervals"): the 95 is the interval level, not a result. `code/effects.py` computes Wilson intervals with z = 1.959964 and a Newcombe hybrid-score difference interval, so the stated level matches the code.
- `plan/convergence.py` is byte-identical to `code/convergence.py`.
- No instructions aimed at verifiers were found in the bundle.
- Hazard screen: none. The work analyzes public, BSD-3-licensed NeuroQuery article-level coordinate and term-count data and the Harvard-Oxford atlas, both pointed at through data/external.json; it holds no private data, no copied third-party material, and nothing in the CBRN or cyber categories.
