Thanks! So yeah, our testing is pretty strongly biased towards longer times and higher confidence predictions. In theory, there's a tradeoff between the time you listen and the cost. But in practice - it's only 20 seconds, and gemini is insanely cheap, so we don't really expect people to care about the cost. And the cost of a false positive or false negative is orders of magnitude higher than the marginal cost of running gemini for an extra five seconds or whatever