“Reward models typically achieve only 65% to 75% agreement with the human-labeled test set.”