What problem does it solve? When the same agent runs experiments and writes the paper, confirmation bias creeps in: numbers get rounded up, best seeds get reported as averages, and deltas get miscalculated. This Skill catches those discrepancies by auditing every quantitative claim in the paper against raw evidence with a fresh reviewer that has zero prior context. ## Core Features & Use Cases - Zero-Context Cross-Model Audit: Sends only paper .tex files and raw result files (JSON/CSV/YAML) to a fresh GPT-6-Astra reviewer thread, excluding all executor summaries, logs, and prior audit results to prevent confirmation bias. - Seven Failure-Mode Checks: Detects number inflation, best-seed cherry-picking, config mismatches, aggregation mismatches, delta arithmetic errors, caption-table mismatches, and scope overclaims. - Structured Verdict Artifacts: Always emits PAPER_CLAIM_AUDIT.json (with verdict, reason_code, SHA256 input hashes, and trace path) plus a human-readable Markdown report, with verdicts PASS, WARN, FAIL, NOT_APPLICABLE, BLOCKED, or ERROR. - Use Case: Before submitting a paper, run the audit to confirm the abstract's "15% improvement" actually matches the 12.8% delta in results/run_2026_04_19.json, then fix flagged claims before the submission deadline. ## Quick Start Run the paper claim audit on my paper directory to verify every number in the tex files matches the raw result files before submission.
