Case context: A roadside-sign system first finds a sign in a photo and then classifies the sign type. On one dev example, a stop sign is labeled "no sign" (y=0). The engineer is unsure whether the locating step or the sign-type classifier caused the mistake, because the pipeline only returns the final label.

Question: How should the engineer use the perfect input test to identify which stage needs attention, and what does each possible result mean?

Sample answer: The engineer should create a hand-marked crop that contains exactly the stop sign, then send that crop directly to the sign-type classifier and skip the locator’s output for this test. If the classifier still predicts y=0 with the clean crop, then the classifier is failing on that example and should be the focus of improvement. If the classifier predicts y=1 when given the clean crop, then the classifier is working and the original failure came from the locating step, which must have produced an inadequate region.

Key points:

• Manually specify the correct region for the failing example
• Feed that corrected crop to the classifier by itself
• Wrong even with the clean crop -> classifier is the problem
• Correct with the clean crop -> locator output caused the error

Rubric: Full credit: correctly explains replacing the locator output with a hand-marked region, running the classifier on that input, and interpreting both outcomes correctly. Partial credit: describes the test but reverses one conclusion. No credit: suggests a diagnostic method that does not test the pipeline stage directly.