Handling adversarial errors — LessWrong