Case context: A team building a cat photo application uses classification accuracy as their evaluation metric. Classifier A scores higher on this metric than classifier B. When the team tries out both algorithms on real user images, they discover that classifier A occasionally lets pornographic images slip through, while classifier B does not exhibit this problem.

Question: As the team lead, what should you decide about which classifier to ship, and what does this decision suggest about how the evaluation metric should be revised?

Sample answer: Despite classifier A's higher accuracy, the team should ship classifier B, since letting pornographic images through creates unacceptable harm to users and the product's reputation. This case suggests the accuracy metric should be revised or supplemented so that it penalizes or disqualifies pornographic image leakage, ensuring the metric reflects the actual product requirement rather than just overall correctness.

Key points:

• Classifier A has higher accuracy but leaks pornographic images
• Classifier B does not have this leakage problem
• The team should avoid shipping classifier A despite its accuracy advantage
• The evaluation metric fails to capture this unacceptable behavior
• The metric should be revised to penalize or disqualify this type of error

Rubric: Full credit requires recommending classifier B (or at least rejecting classifier A) with justification based on the leakage issue, and proposing that the metric be revised to reflect this critical requirement.