verify-claims

Cross-family verification validates experiment claims by auditing them against a foreign model family, ensuring objectivity and reducing self-catch errors. It checks research results for accuracy and reproducibility at various experiment lifecycle stages.

Adversarial auditing prevents self-catch errors by using an independent foreign model family as a verifier. This cross-family approach ensures objectivity, reducing the risk of confidently wrong numbers and improving overall research integrity.

Run the verify_claim.sh script with your claim file and evidence directory. This script executes the adversarial verification process using external model runners to audit the experiment results.

You need external model runners like codex and claude, along with the verify_claim.sh script. These dependencies provide the independent model family required to objectively audit and verify experiment claims.

Use cross-family adversarial checks whenever reviewing research experiments to ensure results are reproducible and conclusions are justified. It can be applied at various stages of the experiment lifecycle to validate accuracy.

The best way to ensure research integrity is cross-family adversarial verification, which audits experiment claims against a foreign model family. This independent verification confirms that results are reproducible and conclusions are justified.