# Verifier notes

## Re-run
`code/run` ran in the declared container (python:3.12-slim, numpy 2.0.2, scipy 1.13.1, matplotlib 3.9.4) with no network. It exited 0 in 1.8 s, and all 15 declared C1 values matched exactly.

## Independent check
- **Spontaneous infarction:** recomputed by hand from the counts 30/456 versus 11/454, using a log risk ratio with a Wald interval. The result is 2.72 (1.38 to 5.35), the same as `D.mi_spontaneous`.
- **Published hazard ratios:** `D.death_hr` (1.03, 0.62 to 1.72), `D.repeat_hr` (2.17, 1.40 to 3.36), `D.hard_hr` (1.39, 0.97 to 2.00) and `D.scai_hr` (0.69, 0.50 to 0.94) are the same as the FAME-3 stable-subgroup rows in `data/hr_effects.csv`. That file cites doi:10.1016/S0140-6736(25)00505-7, Appendix Table S9. I did not check the source table itself.

## Finding on what the computation establishes
`code/analyze.py` line 144 writes several declared values into the results dictionary `D` as string literals rather than deriving them from `data/`:
- `n_stable` 1013;
- `death_hr` and `death_hr_ci`;
- `repeat_hr` and `repeat_hr_ci`;
- `hard_hr`, `scai_hr` and their intervals;
- `spont_counts` and `repeat_counts`.

The hazard-ratio literals equal the values in `data/hr_effects.csv`. For these items, reproduction confirms that the code outputs the transcribed published values. It does not confirm that they were extracted from the source correctly; that rests on the source table.

The values computed from data are:
- `mi_spontaneous.rr` and its interval (recomputed above);
- `definition_shift_pp`;
- `mortality_excludes_20pct`, which is false because the interval's upper end, 1.72, exceeds 1.20.

## Hazard screen
- The bundle is an aggregate-data synthesis of published trial results.
- It contains no patient-level records and no credentials.
- It contains no copied text or tables beyond cited numbers.
- I found no instructions aimed at verifiers.

Answer: none.
