It did, no? If you send a JSON schema to an LLM endpoint, at least for OpenAI and Anthropic, they turn it into a grammar behind the scenes and use it for constrained decoding.