On Mon, 27 Feb 2012 11:48:50 +0100 Simon Loewenthal wrote:
> Recently I enabled shortcircuit for ham on a server because Bayes > seems reasonably well trained. It works well. > > I noticed that emails that did not hit BAYES_00 (so no shortcircuit) > were not autolearnt by SA. Even though these were well below the > autolearn threshold of -1. > > In the example below, the score was -7.8. Below this, the bayes > settings in the local.cf are listed. A score of beneath -1 should > have been autolearnt. > > No, score=-7.8 required=5.0 tests=DCC_CHECK,RCVD_IN_DNSWL_HI, > RCVD_IN_RP_CERTIFIED,RCVD_IN_RP_SAFE,SPF_PASS,T_RP_MATCHES_RCVD,URI_HEX > shortcircuit=no autolearn=unavailable version=3.3.1 > ... > Ham that shortcircuits has autolearn=disabled, which makes sense to me > as it is already in Bayes. Correctly identified isn't the same as "in Bayes", it's presumably because its not considered safe to autolearn on a incomplete set of rules. autolearn=unavailable means that there was some problem with accessing Bayes - note that there is no BAYES_* hit at all in the test list, it's not just that it lacks BAYES_00. I think this might be due to an auto-expiry. Try setting "bayes_auto_expire 0" and running "sa-learn --force-expire" from cron in the middle of the night. This is considered best-practice in any case.
