John Rudd wrote:
I had a similar problem. I don't divert unknown addresses to salearn,
but if I don't fish a message out of my spam folder within X days, it
gets automatically sent to sa-learn and awl.
Then, last week, I started seeing BAYES_00 on messages that would have
otherwise been scored as spam. I responded by removing the negative
values for low bayes probabilities.
Wait, how does training random text as spam indicators result in Bayes
thinking that text indicates ham?
At worst, I can see it diluting the Bayes scores for strong indicators,
resulting in more hits close to BAYES_50. But to trigger BAYES_00 means
that you have to have trained something similar *as ham*.
I expect this has less to do with automated training (since, by your
description, we're not talking auto-learn, so it won't learn anything in
that folder as ham) and more to do with a new type of spam that
simulates real mail more effectively, or that manages to get
auto-learned in the initial SA process (if you have auto-learn enabled).
--
Kelson Vibber
SpeedGate Communications <www.speed.net>