A machine-learning preprint reports higher held-out likelihood from correlated uncertainty modeling, with limits in scale, stability and scope.