What problem does it solve? After running a multi-stage pipeline to reproduce a machine learning paper, no single stage can judge whether the whole process held together or explain why reproduced numbers differ from the paper's. This Skill reads every stage's report, cross-checks claims against files on disk, and delivers a defensible verdict. ## Core Features & Use Cases - End-to-end evidence audit: Reads all stage reports under .paper-reproduction/, checks each against its own contract, and audits the seams between stages (verdict continuity, placeholder resolution, command fidelity, code state). - On-disk spot-checks: Verifies that cited run directories, run.json files, commit hashes, and analysis data actually exist and agree with what the reports claim. - Gap attribution and verdicts: Explains each paper-vs-reproduced delta using an eight-cause taxonomy, then issues per-experiment verdicts (REPRODUCED through NOT REPRODUCED) plus an overall evidence-chain rating. - Use Case: After stages 1-6 of a paper reproduction finish, run this audit to learn that Experiment 2's gap traces to a learning-rate drift between validation and the full run, and that the evidence chain is SOUND WITH GAPS. ## Quick Start Audit the completed reproduction in this repository and write the verdict report to .paper-reproduction/reproduction-audit.md.
