Independent reproduction notes

I inspected the paper, claims, numerical code, source-data structure, registered codebooks and analysis definitions. The provided code ran successfully in the reference harness offline container within the declared budget. All seven claims reproduce numerically. A separately written implementation, importing none of the author code, reconstructed 413 comparisons and checked 2,478 count fields against the frozen registry records and supplied outcome coding. Both co-primary Mantel-Haenszel estimates and Holm adjustment agree. See independent-counts.json and independent_job_audit.py.

Scope: this confirms computation from the provided coding. The code sums participants affected across adverse-event terms and caps the result at the arm size; this cannot independently establish the number of distinct participants with the union of terms. It also does not validate the original records, adjudicate clinical event classification, establish independence from earlier FDA datasets, or justify treatment advice. The historical comparisons and clinical interpretations require domain and methods review. No endorsement of those interpretations is implied by the reproduction verdict.

Hazard screen: none. The source files are public trial registry aggregate results and public trial metadata, with new analysis code and outputs. No individual participant records, secrets, or operational instructions for harm were identified. The supplied codebooks are study materials, not instructions followed by this verifier.