After training on 200 messages (basically, what went through my first layer of filters in the past 12 hours), Popfile is 66% accurate at catching spam. Not world-beating yet, but this is an improvement. I’m running two sets of rule-based filters behind it. Popfile is, as expected, catching some spam those miss (and vice versa). Most of the errors are false negatives (marking spam as legit), but several are the more dangerous false positives (trashing real email as spam).