While I'm waiting to hear back on the bug tracker about whether some odd binascii behaviour is intentional, I was doing some reading. Some of the work by these guys http://www.cs.washington.edu/research/progsys/wasp/wasp.html is very interesting, especially their latest Automated Soundness Proofs for Dataflow Analyses and Transformations via Local Rules http://www.cs.washington.edu/research/projects/cecil/pubs/popl05.html and Automatically Proving the Correctness of Compiler Optimizations http://www.cs.washington.edu/research/projects/cecil/pubs/pldi03.html (There's already been mention on the list of another of their alumni, Grove's thesis on interprocedural optimization.) BTW, structmodule is complete in the sense it passes test_struct.py on my box. Unfortunately (as mentioned earlier) since I don't know of any way to get access to certain C information it fails when I turn off my HAVE_NATIVE flag indicating the hardcoded values are correct.. including, frustratingly, for the default case. There are also some 754 issues involving NaNs I just threw up my hands at, though I have a hack for some infs. I've also discovered that it's relatively easy to get the behaviours for well-formed inputs to match when the algorithm is well-specified, harder to match half-specified behaviour, and brutal to get the errors to match.. like an idiot I tried to implement this one algorithm from the RFC spec I thought it was respecting. Was that ever a mistake.. Doug -- Queen's University Astronomy Research Group Planetary Dynamics Division "We create worlds."