A technique that restricts LLM token generation to follow a predefined grammar, preventing invalid or hallucinated outputs.