On my current corpus, 3.0svn is creaming 2.6-cvs and we haven't even run the score optimizer since 2.6.0 was released (so the 3.0-svn results are orders of magnitude worse than ultimately expected since tons of rules are getting the default score of 1.0).
2.6x branch: ham FPs 73 / 14980 spam FNs 2888 / 14999 3.0 branch: ham FPs 22 / 14980 spam FNs 1979 / 14999 No net, no bayes. I'd build a Debian package and install it right now on my mail server, but the Debian packaging files disappeared at some point. :-( Daniel
