Our assistant passes single-question evals but users say it falls apart over a conversation. How do you evaluate multi-turn quality?

Make your call first. Use the published answer and rubric to examine your reasoning when you are ready.