Version history and performance trajectory for each prediction model. Lower log-loss = better calibration.