Build AI evaluators grounded in real failures.
Recent revisions are kept. Restoring one creates a new revision.
