Question: In a two-stage workflow where a receipt image is first converted to text and then a classifier decides whether the expense is reimbursable, explain why replacing the first stage’s output with a manually verified perfect transcript can localize an error to one stage. Describe what each outcome tells you.

Sample answer: By feeding the second stage a perfect transcript, you change only the quality of the input coming from the first stage while leaving the second stage itself unchanged. If the classifier still gives the wrong result, then the mistake cannot be explained by a bad transcription, so the failure must be in the classifier. If the classifier now gives the correct result, then it was capable of solving the task all along and the original error came from the transcription stage. This experiment is useful because it holds one part of the pipeline fixed and changes only the suspected upstream source of error.

Key points:

• The second stage is tested with an ideal input from the first stage
• A failure that remains points to the second stage
• A success after substitution points to the first stage as the original cause
• The comparison isolates the effect of the upstream stage

Rubric: Full credit: explains that substituting a perfect upstream output isolates the upstream stage’s influence, and correctly interprets both possible outcomes as evidence for either the downstream stage or the upstream stage. Partial credit: describes the procedure but not the causal logic. No credit: does not connect the substitution to error attribution.