Fine-tuning bayes

[Michael Vezie <ml...@ll.mit.edu>]

I'm using SpamAssassin 3.1.7  as an individual user.  I carefully 
sa-learn all spam (and all ham with a bayes score over 0), and it's 
rather good now, but it's not perfect.  I have an idea, and I'd like to 
bounce it off the community to see if it makes sense.

I have various filters in procmail to move spam to different folders, 
based on what address it was sent to.

I set up a filter in procmail to move all spam that doesn't need the 
bayes score to be classified as spam (the score is over 16, for example) 
to it's own filter, and stopped sa-learning it as spam (I just delete 
it).  The idea being, that will free up the bayes filtering for more of 
the gray zone spam. 

Does that idea make any sense?  I've already noticed some of the high 
score spam getting lower bayes scores.  But I'm not sure if it will 
increase the bayes hits of the others.

One other thing I did is I changed the score for BAYES_99 to 9.0, all 
but guaranteeing that anything that the bayes thinks is spam really is.

Michael