Auditing games for high-level interpretability — LessWrong