A message was scored with bayes only and it gave it 0.1. So it was auto-learned as ham. Then I did an sa-learn --spam on it. I re-ran spamassassin on it to see the value. Bayes gave it a 5.1 score (I changed it from the default value, I believe) and when AWL came in, it put it somewhere in the middle (2.x or 3.x). So I ran it through a bunch of times and eventually it averaged out to > 5. So when bayes missed it, it cause the ripple effect, whereas if bayes had known it was highly spam, it would have given a higher score and AWL wouldn't have classified it as ham.
That's a legitimate problem where you might want to manually tweak the AWL using the command line. I'd really just use the remove feature on it in that case and let it rebuild a new average from scratch.
I merely worded caution strongly because there's a lot of admins that get scared when they see the AWL score -5.0, and the final email score is 21.3. That kind of thing is quite normal, and healthy, and just means that SA thought the email should have scored 16.3 (which is still spam) and split the difference between 16.3 and 26.3.
