It's a quantile. Adaptive thresholds are standard — scale τ by the user's own distance distribution. It's a quantile, not a research project.
Two populations, one dial. We've seen this in tag-suggestion features: one threshold, two populations, one of them furious. It never converges.
