I have this account's passphrase back (it was obvious when I saw it, but then these things always are like that I guess). Thanks to Telsa and to yosh for their help.
I've been wondering about the way that the current group of probabilistic spam-filters, from Vipul's Razor via spamassassin to those inspired by Paul Graham's work, actually collect their spam/non-spam corpuses, and, where appropriate, adapt their n-gram and other lexical analyses. I'm putting that here in order to embarrass myself into writing something about it in the very near future.