> what occured to me later is that when MD sends a message to spamassassin
> to analyse, it hasn't modified it. If that message scores more that the
> auto_learn_threshold, then sa will auto learn the original, right?
This might cause problems, depending on how "original" things are. I think
MD needs to put a valid Received header in there, or you may have probems
with the trust zone and end up triggering ALL_TRUSTED for a lot of things
that shouldn't.
> there's another question - if i feed a system-wide bayes ham from only
> one user, will that cause wierdness?
Depends. If the one user has ham that is pretty much like everyone else's
then you will probably be ok. But if other people get other kinds of things
then you could end up getting FPs and possibly FNs for other people.
Bayes is only as good as the data it is fed. Ideally it should have a
fairly representative sample of ham for all of the users to do a good job,
when using site-wide bayes.
If you have users using Outlook or OE you coudl set up some shared folders
that were IMAP folders, and let users drop spam and ham into them, then scan
it yourself (for people dropping the wrong stuff in the wrong folders) and
feed it to sa-learn.
Loren
