Optimal predictor schemes pass a Benford test — LessWrong