We reduced voice agent hallucination rates from 4-5% to less than 1% in production without adding latency. LLMs generate text faster than humans can speak it. That speed gap is where we run detection....