+I replied that "obviously a man" is unsophisticated. I had been thinking of gendering in terms of [naïve Bayes models](https://www.lesswrong.com/posts/gDWvLicHhcMfGmwaK/conditional-independence-and-naive-bayes): you observe some features, use those to assign (probabilities of) category membership, and then use category membership to make predictions about whatever other features you might care about but can't immediately observe. Sure, it's possible for an attempted clocking to be mistaken, and you can have gender categories such that AGP trans women aren't "men", but they're still not drawn from anything close to the same distribution as cis women.
+
+She replied with an information-theoretic analysis of passing (which I would [later adapt into a guest post](/2018/Oct/the-information-theory-of-passing/)). If the base rate of AGP transsexualism in Portland was 0.1%, someone would need log<sub>2</sub>(99.9%/0.1%) ≈ 9.96 ≈ 10 bits of evidence to clock her as trans. Thus, the prospect of passing in naturalistic settings is a different question from whether there exists evidence that a trans person is trans. There _is_ evidence—but who cares, as long as it's comfortably under 10 bits?
+
+I agreed that for most people in most everyday situations it probably didn't matter. _I_ cared because I was a computational philosophy of gender nerd, I said, [linking to a program I had written](https://github.com/zackmdavis/Persongen/blob/8fc03d3173/src/main.rs) to simulate sex classification based on personality, using data from [a paper about sex differences in the "facets" underlying the Big Five personality traits](http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3149680/). Sophia was impressed, but had some cutting methodological critiques. The paper had given the residuals of each facet against the other, so I assumed you could sample one, and then use the residual stats to get a "diff" from one to the other. Sophia pointed out that you can't actually use residuals for sampling like that, because the actual distribution of the residual was highly dependent on the first facet. Given an unusually high value for one facet, taking the overall residual stats as independent would imply that the other facet was equally likely to be higher or lower, which was absurd.