@samth You wrote that Sham “scales to a wide variety of DSLs: some need only a collection of functions, some need only code generation, and some need a comprehensive approach. We present examples of all three as case studies in this paper.” What mapping from examples to kinds do you have in mind?

I’m going to be busy on Monday, except perhaps at 11am. Just to check in, does it make sense for the three of us to meet some time today (Friday), like between 10:30am and 1pm?

I just made a first pass through the Hakaru section. Here are the things that revising it made me want from Rajan: • Please point me to where the Hakaru IR AST is defined in SCF, so that I might excerpt it • Please point me to where the Hakaru IR AST is transformed using map generic, so that I might excerpt it • We discussed measuring Naive Bayes performance on the Sham backend, the old Haskell backend, and MALLET, but it’s also worth repeating the ablation study on Naive Bayes. (For this paper, the ablation study does not need to include omitting the histogram optimization.)