01
1. Pick two runs
- Open /runs and copy the run_ids of the baseline and the candidate (typically pre-deploy and post-deploy).
- Navigate to /runs/compare?ids=A,B. The page renders score deltas, MITRE coverage diff, and finding lifecycle.
- If the inputs were on different templates, the page warns — comparisons are most useful when both runs ran the same scenario set.
02
2. Read the verdict
- Look at brittleness delta first — a literal-coverage gain that loses paraphrase coverage means the fix is fragile.
- Findings tagged new in B are gaps the hardening opened or unmasked. Resolved in A are wins.
- Export the diff as part of an evidence pack so the change history is replayable in audit.