khafra comments on It’s hard to make scheming evals look realistic for LLMs

[khafra]

Linear probes give better results than text output for quantitative predictions in economics. They’d likely give a better calibrated probability here, too.