Don’t A/B Test the LLM: Randomize the Voice Conversation Policy

A production-ready experimentation architecture for AI voice companions that preserves randomization across turn-taking, interruption, latency, and re