Test language models for bias and privacy leaks.
Recent revisions are kept. Restoring one creates a new revision.