Were I to design it, I would pagerank on users and their votes, or use a bayesian filter on article text with classification via some top users' votes. But I'm very curious to find out pg's solution.