I use a similar setup. Spam is simply discarded, whereas mails from all whitelisted users will force autolearn. Any mails thrown into the spam folder will be pa

I use a similar setup. Spam is simply discarded, whereas mails from all whitelisted users will force autolearn. Any mails thrown into the spam folder will be passed on to sa-learn as spam (and submitted to SpamCop and Razor2). So far, no users have complained about false positives being discarded, and moving mails to the spam folder usually results in similar mails being blocked in the future. The Bayes filter is very, very effective on it's own, but as you mentioned it's not perfect. I've found that using Razor2 as a compliment is very effective.