How do you test turn-taking in a voice agent interface?

With one rubric row graded on real exchanges, not with a slider and a feel. Leonard Jung's account of Sierra's Mossy (November 9, 2026) shows the shape: of four quoted reactions, three were taste and the sliders handled them, and one was a behavior, 'I can't tell when I'm talking and when the agent is talking,' which needed two new states (speaking, Mossy rotates and its grass moves in different directions; listening, it pulls back and pulses). The post does not say the fix was tested. The row I would write: a person who did not build it can say which state the agent is in, within a second, by eye alone, graded pass or fail on 20 real exchanges with people who were not in the room. Under it, the voice repair rate: how often the person talks over the agent, asks it to repeat, or restarts, read weekly by surface and owned by design. The five states I keep for any AI surface are confident, unsure, refusing, waiting, and wrong; a voice visual adds speaking and listening.