Quiz

1. Distinguish product outcome, model target, and decision action.
2. Why can random splitting overestimate deployment performance?
3. Give four leakage paths other than including the label directly.
4. When can accuracy be actively misleading?
5. Distinguish discrimination, calibration, and threshold selection.
6. Why must the current workflow be measured as a baseline?
7. What does a data fingerprint prove, and what does it not prove?
8. How can feature semantics differ between training and serving?
9. Why is repeated test-set inspection a form of training?
10. What makes a slice legitimate rather than metric fishing?
11. Why does correlation not establish that an intervention will work?
12. Name two reasons a simpler model may be selected despite lower mean score.

Answer closed-book, then support each answer with one artifact or counterexample from your lab.