Аудит точности данных в научной статье

★ 7.1 · testing

paper-claim-audit is a Claude Code skill that performs zero-context verification of every number, comparison, and scope statement in a scientific paper by matching .tex source files against raw result files. Unlike the related `/experiment-audit` and `/result-to-claim` skills, it answers the specific question: does the paper report the data truthfully and precisely? A fresh reviewer is invoked via `mcp__codex__codex` (always a new thread, never a reply) using gpt-5.6-sol at ultra reasoning effort, with no access to experiment logs, summaries, or conversation history. The audit checks seven failure modes: number inflation, best-seed cherry-picking, config mismatches between compared methods, aggregation count discrepancies, delta arithmetic errors, caption-content mismatches, and scope overclaims. Designed for researchers who need paper-to-evidence fidelity before submission or peer review.