Don't align agents to evaluations of plans — LessWrong