A Normative Rule for Decision-Changing Metrics — AI Alignment Forum