[EMAIL PROTECTED] wrote: > Hello, > > Since upgrading to 3.14, when I turn on bayes auto-learn with: > > bayes_auto_learn 1 > > and I set the learn boundaries with: > > bayes_auto_learn_threshold_nonspam -3.5 > bayes_auto_learn_threshold_spam 15.5 > > I get unexpected auto-learning. Example: I just saw a spam come > through that scored 9.9, which is enough for it to be tagged as spam, > but it should not be auto-learned as spam. But, in the header it > clearly reads: > > X-Spam-Status: > Yes, score=9.9 required=5.0 tests=AWL,BAYES_99, > DATE_IN_PAST_03_06,DCC_CHECK,DIGEST_MULTIPLE,HTML_40_50,HTML_MESSAGE, > MIME_HTML_ONLY,RAZOR2_CHECK,RCVD_IN_WHOIS_INVALID autolearn=spam > version=3.1.4 > > > Any ideas? SA does not autolearn based on the final message score. So, toss the 9.9 out the window. That's not the number SA compares to the 15.5.
For learning SA uses what the message score would have been if: 1) the AWL is off. 2) Bayes was disabled, including shifting what scoreset is used for all the other rules. 3) all white/blacklists are disabled. This is often *quite* different from the final score. However, in this case I don't entirely understand... The default SA 3.1 scores are: score DATE_IN_PAST_03_06 0.736 0 1.122 0.478 score DCC_CHECK 0 1.37 0 2.17 score DIGEST_MULTIPLE 0 0.233 0 0.765 score HTML_40_50 0.611 0 0.497 0.496 score HTML_MESSAGE 0.001 score MIME_HTML_ONLY 0.414 0.001 0.389 0.001 score RAZOR2_CHECK 0 0.5 0 0.5 score RCVD_IN_WHOIS_INVALID 0 2.151 0 2.234 Adding the set1 scores up, the learning score should have been 4.753. Have you modified any rule scores?