Аудит воспроизводимости статистических результатов

research

audit-reproducibility is a Claude Code skill that cross-checks every numeric claim in a manuscript — point estimates, standard errors, p-values, counts, and percentages — against the actual outputs produced by an R, Stata, or Python analysis pipeline, reporting PASS / FAIL / EXPLAINED / UNMATCHED per claim. Tolerance thresholds are read from replication-protocol.md; the skill parses manuscripts in .tex, .qmd, .md, and .pdf formats, then locates computed values in .rds objects, .csv summaries, .json files, Stata .log files, and LaTeX table environments. It is built for researchers who need to catch "updated the code but forgot to update Table 2" bugs before journal submission, before releasing a replication package on openICPSR or a journal archive, or after a major revision. Extracted claims and audit results are written to quality_reports/ as JSON for transparent manual review.