Hmm. I enjoy quite good results with Vipul's Razor, although to be fair it does occasionally trash a DSN or three (could be explained by an included spam attached as part of the report). I let Vipul'z Razor (which uses the cloudmark network and was clearly not affected by this issue - it's written in perl, BTW) have first go at incoming mail, and if not classified as spam at once let my bayesian filters go at it, otherwise (spamtrap or no) the mail just gets filed away as almost-certain spam and has the benefit of costing almost no disk I/O. This assumes it got past my SMTP-level defenses, of course. A spamtrap classifies mail as spam and submits to Razor as spam automatically, while a human-assisted process submits most bayesian-classified remaining spam to Razor. The effect is that most mails get noted as spam right away with a low enough false positive rate that I can check quite infrequently, and I can count and check the remaining classified spam myself daily as there's not too much of it for verification with a much greater confidence in accuracy (my database is quite large). Most of the spam actually comes from forwarders, of course, since spambots are mostly caught at the SMTP stage. This includes easy-to-fingerprint 419 scams and such sent from ISP nets and freemails. Oh, and there's near-zero spam in my inbox with no possibility of any mail actually getting lost and a much greater possibility that I will lose patience getting complacent about how damned good my results are and just throwing my spam folders away. Which is all very pleasant. :-)
Cheers,
Sabahattin