On Wed, 15 Aug 2012 17:05:00 -0400 Kevin A. McGrail wrote: > On 8/15/2012 5:00 PM, John Hardin wrote: > > > > Right. It might be prudent to review the defaults before the next > > major release. > I wonder if we shouldn't disable auto-learning by default (assuming > it's on by default)... > > Bayes should really be trained.
It seems to me that bug 6344 from 2010 has some merit. (I was about to file something similar myself.) https://issues.apache.org/SpamAssassin/show_bug.cgi?id=6344 This suggests that lists like RCVD_IN_DNSWL_* should be marked as noautolearn so that when they fail they don't screw-up autolearning- which is what appears to have happened here. This is exacerbated by the fact that autolearning wont learn against a strong Bayes result (quite rightly), so damage can become permanent.
