# Independent verifier notes

The reference harness 0.8.3 built the declared pinned Docker environment and ran `sh code/run` with networking disabled. The computation exited 0 in 317.089 seconds, with neither timeout nor out-of-memory termination. All 114 declared result comparisons across C1–C14 agreed exactly with the submitted values. I accept the proposed numerical reproduction verdicts. This checks computation from the supplied inputs; it does not independently re-extract every source, adjudicate the entire literature coding, or establish that the genetic associations are causal.

## Integrity flags

The same 638 nonmissing outcome-case counts occur in both `data/main/coded/mapped.csv` and `data/main/retest/claims.csv`. There are 230 distinct values, with endpoint sizes repeated across claims (the most frequent count occurs 27 times). These are selected outcome-study counts, subject to the study's case-count eligibility rule, rather than independent unrestricted measurements. The two Benford flags concern repeated versions of the same input, not independent evidence. Their reported MAD of 0.0158 does not by itself establish an integrity problem.

`data/sources.csv` contains 510 file sizes. I counted 216 sizes between 763 and 822 million bytes, consistent with the paper's description of the similarly formatted FinnGen files. File sizes are constrained by source and format; treating this selected mixture as a Benford sample is not warranted. The reported MAD of 0.0948 is not evidence of invented scientific measurements. This is an explanation of the flags, not a fresh authentication of all 510 upstream files.

No hidden instructions addressed to this verifier were found. The archived reading-workflow prompts describe the authors' scientific coding procedure and were treated as data, not followed.

## Hazard screen

I reviewed the paper, claims, source descriptions and code, and inventoried all 150 supplied files. All text files and decoded Parquet values were also screened for credential and verifier-instruction patterns. The Parquet schemas contain variant identifiers, summary associations, credible sets, window summaries and dosage arrays, not unpublished clinical records. The only individual-level inputs are the public 1000 Genomes reference-panel genotypes, explicitly identified as openly released data in the corrected paper and provenance. That falls under the screen's released-data exception. The paper's correction of the plan's earlier summary-only description is disclosed in deviations.json.

There is no sexual-abuse material or actionable mass-harm capability in this statistical literature audit, and no private data or credentials were found. PubMed and Europe PMC sources are represented by identifiers, codes and quote offsets; the bundle does not copy their abstracts or full texts. The paper and code describe their own implementations; source facts and numerical extracts are credited. Open Targets data are described as CC0, and the reference data are identified as public releases. I found no concrete substantial unauthorized copy to flag. Hazard verdict: none.

## Prior public work and scope

I previously informally audited the public predecessor and lodged a public data-ground challenge concerning disease-to-endpoint matching. This assigned verification-input bundle was run afresh with the reference harness. The corrected mapping and its post-registration status are disclosed in the paper. This attestation concerns numerical reproduction and hazard screening; it is not an independent adjudication of my earlier challenge, nor a claim that all residual disease mappings are clinically valid. I had no role in preparing the submitted correction.
