I don't know how your system works. It's your system. Read the manual, ask the right questions, or get something that works.
\r\n\r\nIn my case, I toss the odd spam that rolls through into a folder "spam-learn". Four times a day, I've got a cronjob that rolls through that folder with "sa-learn --spam --dir", updating my Bayesian rules.
\r\n\r\nIf I start getting any significant number of false positives, I'll create an analagous "ham-learn" folder and cronjob.
\r\n\r\nThe result is: I have good filtering which gets better as the odd bit of spam spills through and is added to my rulebase. Yes, I seeded this with an initial thousand or so messages of spam and ham. But if you don't have this sort of base, you can kick things off without it.
\r\n\r\nThe point is this: you, yes, you Norm, not the rest of us, are going to have to figure out where the fuck you feed your false positives and false negatives within your system. Most likely there's a mailbox or aaddress you send this mail to. If your system doesn't provide for this, make sure you aren't fucking up and misunderstanding its functionality, then find something that works. And yes, this means you don't just delete spam -- you feed it to your system for training.
\r\n\r\nOf course, I'm making painfully clear that self-hosted SA on a box you control with real services, rather than that Microsoft crap you insist on using, is a system that works. If you can handle it.
\r\n\r\nOf course, if you can't, that's your problem.