A probability model has good ranking but overconfident probabilities. Which assessment should be added?

1. Only an ROC curve with no probability analysis
2. Only the number of prediction requests
3. Calibration analysis against observed outcome frequencies
4. Only exact label accuracy at one threshold