Your prompt is guarded before the LLM sees it, and every sentence of the response is monitored as it streams — with zero added latency on the happy path.