A rare-event detector labels almost everything negative and obtains 99% accuracy. Which metrics should be examined?

1. Only the total count of negative predictions
2. Only raw image resolution
3. Precision, recall and the confusion matrix for the rare class
4. Only the number of model layers