(Another post-hoc writeup)
I was at the SPARC planning workshop since 03-06 and will be until 03-10, but came back today to give a short talk at lab meeting with updates on my research. Basically, I now have a way to deal with valid abstractions by looking at sub-polytopes of the probability simplex, and predicting using the maximum entropy distribution (which is the Nash equilibrium in a certain game against nature). However, I still need to work out conditioning (I seem to recall thinking a bit about this already, and e-mail logs indicate I did indeed do so on March 4th or perhaps March 5th, and concluded that the worst-case was going to not be good). Current hack is to use the maximum-entropy distribution.