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.
Preserve
Section titled “Preserve”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 Do not publish only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.
Compare
Section titled “Compare”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.