Is this a good idea?

[Thomas Cameron]

Howdy -

My scenario:

SA 3.0.0 and spamass-milter CVS on a Fedora Core box.  It's the only host 
listed as an MX for the domain, but the customer actually uses a large ISP 
for e-mail.  We just have the FC2 box relay all mail for the domain to the 
ISP.  That means that when the customer receives the e-mail it has headers 
for the ISP.  I'm not sure if that's important for what I want to try:

I am thinking of a system using spamass-milter and SA 3.0.0 whereby 
exceptionally spammy (i.e. score over 8 or 10) gets silently copied to a 
spam mailbox on the relay server and exceptionally hammy mail (i.e. score of 
less than 1) gets silently copied to a ham mailbox.  Then sa-learn could be 
run against the mailboxes to feed the Bayes database.

Is that a good idea?  I kind of think that the downside is that a big 
benefit of Bayes is that false negatives can be fed to it so that they are 
more likely to be caught next time.  In the scenario above I would think 
that it's really just reinforcing what is already known.

Past that, I am really dying to figure out a good way to use Bayes on a 
relay server.  I keep hearing that you need to be able to forward messages 
as an attachment but nothing past that.  Any comments, suggestions, etc. 
gratefully accepted.
--
Thomas Cameron, RHCE, CNE, MCSE, MCT