Models based on measures of fatigue and performance make the implicit assumption that they will be successful in predicting risk. The present paper reviews the available literature on shiftwork safety in which real measures of accidents or injuries could be pinpointed in time and in which the a priori risk appeared to be constant. Three main problems for the models emerged from this review: 1) risk was significantly higher on the afternoon shift than on the morning shift; 2) the dominant peak in risk over the course of the night shift occurred at about midnight; and 3) risk increased substantially over spans of four successive nights. It is suggested that the relationship between risk and fatigue may be non-linear, that models may have overestimated the recovery during short sleeps, and that day sleeps between night shifts may be less recuperative than normally timed night sleeps of the same length.