Bob Proulx wrote: > 0.0 RCVD_FAKE_HELO_DOTCOM Received contains a faked HELO hostname > 2.2 FORGED_YAHOO_RCVD 'From' yahoo.com does not match 'Received' > headers > score RCVD_FAKE_HELO_DOTCOM 0.899 0.034 0.969 0.424
Thanks to jdow and also to Bob Menschel who mailed me offlist I now have the answer. My bayes database was (temporarily) broken when that message was scored. grep -r score.*FORGED_YAHOO_RCVD /usr/share/spamassassin score FORGED_YAHOO_RCVD 1.668 2.174 2.095 2.700 Since FORGED_YAHOO_RCVD lists here as 2.2 I was operating on column two. I was expecting to be operating in column four. Column four would print 0.4 but column two would print 0.0 within the rounding and truncation of the formatting. Quoting the docs: If four valid scores are listed, then the score that is used depends on how SpamAssassin is being used. The first score is used when both Bayes and network tests are disabled. The second score is used when Bayes is disabled, but network tests are enabled. The third score is used when Bayes is enabled and network tests are disabled. The fourth score is used when Bayes is enabled and network tests are enabled. Basically: score tag -bayes-net -bayes+net +bayes-net +bayes+net Why was I operating in "-bayes+net" mode? Not sure. I am normally using bayes and getting column four scores. I had just done the upgrade to sa-3.0 and that changes the database from version 2 to version 3. I also did some more training on spam. But my sa-learn --dump numbers were very low in the 250 range even though I should have had a thousand messages trained. I think the bayes db was lost during my upgrade. Then I trained on top of it. In the middle of that this message was scored without bayes but with net and so got the column two scores. That explains the strange behavior. To double check this I recovered my previous bayes db from backup. I downgraded back to the previous sa-2.64. Then upgraded again to sa-3.0 to try to recreate my problem with the bayes database. But I was unable to recreate the problem. Everything worked fine this time and my bayes database lists 1137/1272 ham/spam now. So it must have been some strange problem between the chair and keyboard when I did the first upgrade. Bob