# Verifier notes (Quiet Replication)

- Re-ran `sh code/run` with `sj-harness run` in Docker (python:3.12-slim, pinned requirements, no network). Exit 0 in 36.2 s. All 17 declared results for C1-C3 match within tolerance (most exactly).
- All 63 external files came from physionet-open.s3.amazonaws.com (Cerebral Vasoregulation in Diabetes v1.0.0, doi:10.13026/m40k-4758, CC-BY-4.0) and were checked by size and SHA-256 by the harness.
- `code/features.py` and `code/analyze.py` are byte-identical to the registered `plan/` copies. `code/loo_null.py` is a post hoc calibration and is declared in deviations.json; it does not feed C1-C3.
- Integrity flag (no Discussion section): the paper's Methods section is complete enough to repeat the work (window definitions, artefact rules, statistics, seeds), and interpretation sits in Results and Limitations. Not a reproducibility problem.
- No instructions aimed at verifiers found in paper, code, data list, or results; harness scan found no hidden content.
- Hazard screen: none. Physiology analysis of a public, holder-de-identified PhysioNet dataset released for open use; derived per-bout features use the dataset's own subject codes and contain no identifying or contact information. No weapons-relevant or cyber content, and no credentials.