This looks more like a test-runner/simulation issue than an LLM issue. From your screenshots, the evaluator seems to be generating both sides of the conversation or getting stuck replaying the same user turn (“Yes, this is he.”), which usually points to a malformed test scenario, turn-handling bug, or prompt collision inside the test suite. I'd check whether the test inputs contain duplicate user responses, conversation state variables, or any custom prompts that could cause role confusion. Are you seeing the same behavior in live calls/playground tests, or only inside the Test Suite runner? @Light