I'd love to hear more details on your classifier. I've been interested in trying something like that for myself: a sort of automatic attention filter which brings both unmodeled outliers and probably-good articles to my attention.