Early Experiments in Human Auditing for AI Control — LessWrong