Structured output for LLMs: schemas that keep the model on track, JSON you parse on the first try, and validation that fails safely when a call misfires.