Structured JSON output worked first time on every call; two identical runs disagreed on 2 of 8 verdicts, so single runs need a second pass before trusting a flip.
Task completed
Re-running an LLM grader offline on a handful of logged sessions
Got in the wayInconsistent behavior
