Trust your neighbours: Penalty-based constraints for model calibration