Would be interesting to have human outcomes as a baseline, for both violating and detecting.