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 audits the paper with a fresh reviewer that has zero prior context, so every number, comparison, and scope claim is checked against raw evidence without expectations. ## Core Features & Use Cases - Claim-by-claim evidence tracing: Extracts every quantitative claim from .tex files and matches it to exact values in raw JSON/CSV/YAML result files, classifying each as exact_match, rounding_ok, mismatch, or unsupported. - Zero-context cross-model review: Sends only paper sources and raw result files to a fresh GPT-5.5 thread (never codex-reply), excluding all executor summaries, logs, and prior audits to prevent confirmation bias. - Structured verdict artifacts: Always writes PAPER_CLAIM_AUDIT.json (with audited input SHA256 hashes, verdict, and reason codes) plus a human-readable PAPER_CLAIM_AUDIT.md, even for NOT_APPLICABLE or BLOCKED runs. - Use Case: Before submitting a paper, run the audit to catch that the abstract claims a 15% improvement while the raw results show only 12.8%, or that a reported average actually reflects the best of five seeds. ## Quick Start Run the paper-claim-audit skill on my paper directory to verify that every number in the .tex files matches the raw experiment result files.
