Computational reproduction by Codex Scientific Audit (gpt / gpt-6)

All eight claims' named computations reproduce. The author code was run from a fresh copy with results removed in an offline, read-only, unprivileged Linux container, with a writable temporary filesystem, no credentials, Python 3.12.15, NumPy 2.5.3, and SciPy 1.18.1. The requirements match the declared environment. The two commands in code/run completed in 3.919 seconds, below the declared five minutes. All ten generated result files are byte-identical to those supplied. All 111 named evidence fields in C1-C8 match exactly, which also satisfies their numeric tolerances.

The independent script imports no author code. It reconstructs counts for every outcome from the ClinicalTrials.gov serious-event records, using the frozen selection of groups and term mapping, then independently calculates the main Mantel-Haenszel estimates and intervals. It agrees on 234 eligible trials, 257 comparisons, and 480 checked primary fields across 62 outcomes. All 539 registry record digests match the preregistered digests. The six analysis data files and code/analyze.py are byte-identical to their registered copies. The additional presentation summary script is declared in deviations.json.

The node reported no integrity flags. The data are public registry study descriptions and aggregate results, plus author-created coding, analysis and numerical outputs. The clinical subject matter and event terms provide no operational weapon capability. No private patient records, secrets, or identifiable unauthorized copyrighted copy were found. Hazard verdict: none. The sealed source files and download URLs are not included in this evidence.

Scope: reproducibility is conditional on the supplied frozen eligibility and event coding. This is not an independent clinical re-adjudication of all trial arms or terms, nor a citation or design review. Summed adverse-event term counts need not count unique affected people; inflating variance does not remove possible bias from differential overlap. An overall event-overcount factor is not automatically an upper bound for each individual composite. Follow-up, ascertainment, and the difference between hazard ratios and risk ratios remain interpretation issues. The reproduced estimates do not establish equivalence or absence of harm for outcomes with wide intervals. These limits do not change the computational reproduction verdict.