Aren't llms now outperforming humans in prediction markets? If so, that seems as if they would probably be quite good at confidence scoring.