Judge subjective AI failures with a focused prompt.
Recent revisions are kept. Restoring one creates a new revision.
