On (assessing) the fairness of risk score models