experiment-audit

Detect fake ground truth, score normalization tricks, and phantom results by running cross-model reviews on evaluation scripts, results, and paper claims. The audit enforces artifact provenance checks to verify data traceability and consistency across multiple experiment files.

Cross-model review is a mechanism that routes experiment artifacts to an independent reviewer backend defined by an audit protocol. It verifies provenance and evidence alignment across multiple files to detect fabricated ground truth and phantom results.

Verify reproducibility by analyzing the target experiment directory with standardized artifact collection and cross-model reviews. This generates an EXPERIMENT_AUDIT.md report and a machine-readable JSON file highlighting provenance, scope, and evidence alignment.

Experiment integrity auditing requires a reviewer backend defined by the audit protocol, a standardized file collection, a vulnerability-free workflow, and explicit config files. No external dependencies are needed to execute the audit process.

Cross-model experiment audits require a defined reviewer backend and explicit config files to function. The audit is limited to verifying provenance and consistency across provided artifacts and cannot detect issues outside the submitted experiment directory scope.