Machine learning needs a great deal of training data, usually labeled by people. To find out how well a sorter works, we hold back test cards it never learned from, measure accuracy as right answers out of the total, then refine the sorter and test again on fresh cards.