bturtel

Great question! It's probabilistic so not really "right vs wrong" on any single question, but who better estimated the likelihood. One big difference shows up when there's no useful context - we ran the same eval WITHOUT including any useful up-to-date context with questions. In this case, GPT-5 stays overconfident and its BSS drops to -11.3% (vs -4.3% ours) - worse than just guessing the base rate. So one advantage of the RL training is just learning to know what you don't know, and identify when there's real signal.