One of the interesting things that I have noticed while watching my latest corpus grow is the changes that are made in the score of message that are originally scored to be neutral or unknown.
I think it would be nice to recheck messages on some sort of interval and get new er scores, adjusting only neutral messages to make them either good or bad based on a newer corpus then the one they were originally scored with.
During this process it occurred to me the virus scan has some of the same issues. Most notably that the anti-virus database changes at least once a day and the last thing anyone wants is to have a virus get into their email box.
So along with re-scoring messages I am also rescanning them for viruses as well.
Either way it keep the database a bit cleaner and help reduce the user interaction with their neutral messages. Given enough newer messages in the corpus any and all neutral message will, in time, be classified either good or bad.