I think I've bitten off more than I can chew here... Any filter gurus want to take a crack at this?

I'm using spamassassin in an xfilter command and have suspected spam returned to Maildrop with the subject rewritten like this:

"SPAM xx: message_subject" or "SPAM xx.xx: message_subject"

where "xx" or "xx.xx" is either an integer or decimal number with one or two decimal places and message_subject is... the original subject of the message. The numbers are spamassassin's _HITS_ variable that denotes the apparent spamminess of the message. Spamassassin also adds an X header that is used in a simple Maildrop filter to direct suspected spam to a mailbox.

As I've been using spamassassin I've noticed that a message with a low score (say, under 10) has a slight chance of being a false positive (legitimate message) but a score over 10 is almost certainly spam. I'd like to create a filter that filters low-score spam to my spam mailbox for forensic study but high-score spam is dropped into the infinite void of /dev/null.

Any clues? Thanks!