What problem does it solve? LLM agents can unintentionally produce fraudulent experimental results, such as fabricating ground truth from model outputs, self-normalizing scores to near 1.0, citing nonexistent result files, or overstating evaluation scope. This Skill adds an independent integrity check before any claims are written. ## Core Features & Use Cases - Cross-Model Independent Review: The executor only collects file paths; an external reviewer backend (Codex MCP or Manual Review MCP) reads the code and judges integrity, preventing self-acquittal. - Six-Point Fraud Checklist: Checks ground truth provenance, score normalization, result file existence, dead code detection, scope assessment, and evaluation type classification. - Structured Audit Reports: Writes EXPERIMENT_AUDIT.md and EXPERIMENT_AUDIT.json with PASS/WARN/FAIL verdicts that downstream skills like /result-to-claim and /paper-write can consume. - Use Case: After an experiment pipeline finishes and before writing paper claims, run the audit to verify that reported metrics actually exist in result files and that ground truth comes from the dataset rather than model outputs. ## Quick Start Ask the agent to audit the experiment results in your project directory for integrity before writing any claims, for example by saying "audit results in ./experiments/run-03".
