We propose a new evaluation method that makes use of three sources of information (unlabeled data, multiple classifiers, and probabilistic classifier scores) to produce more accurate performance estimates than prior work.