# Verifier notes

## Hazard screen

Answer: none.

Read paper.md, claims.json, code/ (run, analyze.py, check-r.R), data/, env/, plan/, results/, materials.json, provenance.json and references.json as data. The work is a study-level audit of published aggregate adjusted hazard ratios from seven moderate aortic stenosis cohorts and two trials. No CSAM, no weapons or unauthorized cyber capability, no individual patient records or secrets (only aggregate counts and published estimates), and no substantial copied third-party text, tables or figures identified. No verifier-directed instructions found.

## Run

Ran `sh code/run` through the harness in rootless Podman with no network: exit 0 after 3.1 s (declared 2 min). R1.json, R2.json, R3.json and validation.json produced in the container are byte-identical to the bundle's results/. The script's built-in comparison with the archived R/metafor reference (data/metafor-reference.json) passed with a maximum absolute difference of 1.9e-6 (tolerance 1e-5). The R reference itself was not re-run, since its environment installs CRAN packages at build time and it is not part of code/run.

All 11 C1 values agree with the declared values.

## Independent arithmetic check

Recomputed by hand from data/cohorts.json (Hariri2024 HR 0.51, 0.34-0.77; Adrichem2025 HR 0.51, 0.35-0.76): the inverse-variance common effect gives 0.51 (0.384956-0.675661). Since both point estimates are equal, Q = 0 and tau^2 = 0, so the modified Hartung-Knapp multiplier is floored at 1 and the t(1) critical value 12.706 gives 0.082344-3.158713. Both match. The REML weights for all seven cohorts are Hariri 12.442781 and Adrichem 13.562047, which sum to 26.004828%, matching 26.0048.

## Observations (not affecting the verdict)

- The 72 "delayed valve replacements" figure is derived as 74 deferred valve interventions minus 2 balloon-only procedures (data/audit-counts.json). The claim statement calls it "72 reported delayed valve replacements", but the count the bundle records as reported is 74. The derivation is shown in code.
- The Angellotti percentage (22/116 = 18.9655%) is hard-coded in analyze.py rather than read from audit-counts.json. The values agree with the counts file.
- With only two cohorts, the Hartung-Knapp interval (0.08-3.16) is very wide. The paper says itself that this two-study variance inference is unstable.
