On Mon, Jun 16, 2003 at 04:43:53PM -0400, Don Armstrong wrote: > On Mon, 16 Jun 2003, Duncan Findlay wrote: > > The only negative rules will be: bayesian rules, bondedsender and > > habeas. Figuring how to autolearn ham (non-spam) is the only obstacle > > we still need to figure out. > > Sure sounds like throwing the baby out with the bathwater... but I > presume you all are running statistics on email distributions...
Eventually, spammers will forge any test they can. (This of course presumes that spamassassin is a big problem for spammers.) It's extreme, but necessary. All the spamassassin scores are generated with a genetic algorithm using results from about 150k spam and 150k non-spam. The scores will naturally be adjusted to compensate for the lack of negative scoring rules. Anyways, this is quite OT for debian-devel (although so is the vast majority of this thread....). -- Duncan Findlay
pgpeCXv0A64y5.pgp
Description: PGP signature
