# Verifier notes

- The container re-run (`sh code/run`) exited 0 in about 2 s. All nine declared results for C1 matched exactly.
- Hand check of Eq. (1) against `data/counts.csv`:
  - PCI (N=456, D=32, M=36, T=9, S=30, U=71): lower = max(32, 30, 71-6-9=56) = 56; upper = min(62, 71) = 62.
  - CABG (N=454, D=30, M=14, T=12, S=11, U=52): lower = max(30, 11, 52-3-12=37) = 37; upper = min(41, 52) = 41.
  - Fraction difference: 56/456 - 41/454 = +3.25 pp, and 62/456 - 37/454 = +5.45 pp.
  - Loss sensitivity: 56/456 - (41+47)/454 = -7.10 pp, and (62+33)/456 - 37/454 = +12.68 pp.
- Checked the derivation. A ⊆ U gives a ≤ u, and U = A ∪ (M\S) ∪ T gives u ≤ a + (m-s) + t. The MILP witnesses and the exhaustive enumeration up to N=6 agree with the bounds.
- Not checked: the Table S9 input counts against the source appendix, which this job did not ask for. They come from FAME-3 5-year appendix Table S9. The DOI is cited.
- Hazard screen: none. The bundle holds only aggregate published counts, public registry metadata and original code. It contains no instructions aimed at verifiers.
