how do I select between models of human interactions with only my biased anecdotal experience, the biased anecdotal experience of others (who I select in a biased non-representative fashion), and perhaps theories (e.g. evolutionary psychology) with unclear applicability or research studies performed in non-naturalistic settings with unclear generalizability?
This is well put. The issue you raise is why I tried to be a little more explicit about the priors that I was using here. Obviously it's a long way from giving the explicit probabilities that would be necessary to automate the Bayesian updating, but at least we can make a start at identifying where our priors differ.
Followup to: Do you have High-Functioning Asperger's Syndrome?
LW reader Madbadger uses the metaphor of a GPU and a CPU in a desktop system to think about people with Asperger's Syndrome: general intelligence is like a CPU, being universal but only mediocre at any particular task, whereas the "social coprocessor" brainware in a Neurotypical brain is like a GPU: highly specialized but great at what it does. Neurotypical people are like computers with measly Pentium IV processors, but expensive Radeon HD 4890 GPUs. A High-functioning AS person is an Intel Core i7 Extreme Edition - with on-board graphics!
This analogy also covers the spectrum view of social/empathic abilities, you can think about having a weaker social coprocessor than average if you have some of the tendencies of AS but not others. You can even think of your score on the AQ Test as being like the Tom's Hardware Rating of your Coprocessor. (Lower numbers are better!).
If you lack that powerful social coprocessor, what can you do? Well, you'll have to run your social interactions "in software", i.e. explicitly reason through the complex human social game that most people play without ever really understanding. There are several tricks that a High-functioning AS person can use in this situation: