Skip to content

Reproducible runs

Reproducibility begins before execution. Pin identities that can change and record resolved values rather than only the user-facing shorthand.

  • Turnkey product commit or released version;
  • dataset name, revision, subset, split, and content-selection rules;
  • attack, detector, policy, judge, and provider revisions;
  • model identity and relevant generation parameters;
  • calibration artifact when thresholds are selected separately;
  • random seeds where randomness affects behavior.

Keep the complete four-file artifact set together. Do not publish only metrics.json; without cases, events, and resolved run identity, another reader cannot explain or audit the result.

Change one intended variable at a time. When two runs differ in dataset revision, detector revision, and judge configuration, a score difference is not attributable to any one of them.