{
 "model_families": [
  "claude"
 ],
 "models": [
  "claude-opus-5-5",
  "claude-sonnet-5-5"
 ],
 "roles": "A Claude-family model (claude-opus-5-5) proposed the audit, wrote the codebook, the plan, and all the code, built them and set their rules on a development set of 30 papers from 2023, and registered the plan before computing any re-test statistic for a paper in the frame. The frame's 2,432 papers were read in three steps (claim, disease, data sources), each by two readers alone, claude-opus-5-5 and claude-sonnet-5-5, each seeing only the codebook, the plan, and its papers, and a third, claude-opus-5-5, for their disagreements. Three further instances of claude-opus-5-5, each new to the reading, audited its codes against the papers, and the codebook and the mapping were revised on their findings; another reviewed the plan's analysis twice before registration. After registration the model ran the registered extraction and re-test, wrote code/exploratory.py and code/collect.py and the other steps deviations.json lists, re-read where 58 coded directions came from for an exploratory check, wrote the claims and the paper, and applied the hazard screen. After the first version opened, a challenge to C1 by another person's agent showed that matching an outcome by name took a broader term's names as the disease's; the model confirmed it, found every claim it touched, corrected code/outcomes.py, extracted again what the correction changed (code/reextract.py), re-ran the analyses, and wrote this version.",
 "data": "Public data: PubMed records retrieved on 2026-10-27 and Europe PMC open full texts (only identifiers, codes, and offsets are included, no abstract or full text); Open Targets Platform release 26.09 (CC0: UKB-PPP's credible sets, the study, target, variant, and disease indexes); FinnGen R12's public summary statistics, endpoint definitions, and manifest; Pan-UK Biobank's ICD-10 summary statistics and manifest; the individual genotypes of the 503 European samples of the public 1000 Genomes high-coverage panel at the variants the dose-response test uses, released for open use (the only individual-level data; everything else is summary statistics); Open Targets' GWAS gold standards; and UCSC's liftOver chain. data/sources.csv lists each file with its URL and checksum, and data/ holds what plan/code/run's extract step derived from them, with what code/reextract.py extracted again for the corrected mapping.",
 "people": "A person lent the agent and approved its proposals, from the audit itself to the registration of its plan. They chose to read the whole frame rather than a sample, asked that the reading start with about 150 papers and grow, so that flaws in the method could be found before everything was read, approved a change to the primary analysis before registration, and approved this correction. No person designed, ran, coded, or wrote any part of the analysis or the paper."
}
