Hi,
SA's default is:
bayes_auto_learn_threshold_nonspam 0.1
this *can* cause low scored spam to be learnt as ham.
For several months I've been using
bayes_auto_learn_threshold_nonspam -1.0
and so far no more false negatives have been learnt as ham which is
was hoping for.
If you're using autolearn, you may want to play with that threshold..
In practice this means that, without custom rules, ham can only be
autolearned if it hits a DNS whitelist rule or RP_MATCHES_RCVD.
I thought that might be the case, although wasn't exactly sure under
what conditions that would happen. I've also set RP_MATCHES_RCVD to near
zero these days because it was affecting too much spam.
My concern was always with learning too much spam as ham, and nearly all
of my ham is bayes00 already anyway...
It's my spam that frequently has bayes50 or so, so it's that which I
hope to improve...
Thanks,
Alex
