Case context: You are training a handwritten-digit classifier. Your development set contains 800 labeled images, and the model gets 32 of them wrong. A teammate suggests checking only those 32 wrong predictions to find any bad labels and then leaving the 768 correct predictions alone.

Question: What is the weakness in that plan, and what kind of labeling mistake could still go unnoticed if you follow it?

Sample answer: The weakness is that it ignores the examples the model predicted correctly, and those can still contain bad labels. If an example has the wrong ground-truth label and the model also predicts that same wrong class, the case looks correct even though both the label and the prediction are wrong.

Key points:

• The plan overlooks correctly classified examples.
• A wrong label can hide inside a correct prediction.
• A matching wrong label and wrong prediction can appear to be a true success.

Rubric: The response must explain that only checking the wrong predictions misses some label errors among the correctly classified examples. It must also state that the hidden error is when both the label and the model prediction are wrong in the same way.