experiment-audit

Experiment integrity verification requires auditing ground-truth provenance, metric normalization behavior, and result-file existence against repository evidence. You provide the experiment directory path, and a cross-model reviewer generates EXPERIMENT_AUDIT.md and EXPERIMENT_AUDIT.json with PASS, WARN, or FAIL verdicts.

Metric normalization fraud involves manipulating score normalization to artificially inflate experiment results. Detection is performed by a read-only cross-model reviewer pass that validates ground-truth provenance and checks for dead metrics, phantom results, and over-claimed scope against actual repository artifacts.

Checking for fake ground truth involves auditing whether reported files and scores are genuinely supported by the repository's code and artifacts. The audit process applies path-only artifact collection to verify result-file existence and scope alignment, generating a JSON report with explicit verdicts.

Experiment auditing is non-blocking and will not halt downstream pipeline steps. It produces EXPERIMENT_AUDIT.md and EXPERIMENT_AUDIT.json files and tags unsupported claims as warned, allowing the pipeline to continue while preventing overconfident reporting.

An experiment audit is needed after experiments complete and before making claims in narrative reports or papers. It specifically checks ground-truth provenance, metric normalization behavior, result-file existence, and scope alignment against repository evidence to prevent overconfident reporting.

Yes, experiment integrity auditing uses a read-only cross-model reviewer pass to judge integrity from actual files. The executor collects path-only artifact references, ensuring the validation of ground truth and metrics occurs without modifying the original repository evidence.