hi guys -- thought you might like to know -- I've recently added code to pluginize SpamAssassin's default Bayesian-style probabilistic classifier, allowing other trainable classifier implementations to be switched in in its place.
Given last year's TREC results (congrats ;), the OSBF approach is the first candidate I'm trying out. If you're interested in tracking develoment, the issue tracker is at: http://issues.apache.org/SpamAssassin/show_bug.cgi?id=5686 (The http://www.siefkes.net/ie/winnow-spam.pdf paper has been the main source of data on the algorithms, btw. I suspect it may be a little out-of-date w.r.t. current practices, though...) --j.
