Deployed on an 84,817-image evaluation set, the same previous versions were tested on, equal error rate improved from 3.59% to 3.36% and ROC AUC rose to 99.37%, with one fewer model per request.