“Standard RL training for correctness on the GSM dataset caused human judge accuracy to decrease over time, in contrast to the improvement seen with "checkability training".”