Similar result in production for me. I compared Jev with a general LLM (DeepSeek) as the decider for when to cut live speech into sentences, on the same 226 segments. With Jev the longest held segment was 17 s. With DeepSeek it was 40 s.