Case context: A startup has collected only 48 labeled examples for detecting faults in industrial pumps. One approach is an end-to-end model trained directly from the labels with no built-in domain rules. Another approach breaks the task into stages and includes a module that uses engineers' understanding of common failure patterns.

Question: What should the team conclude about these two approaches, and which source of knowledge should it rely on more heavily?

Sample answer: The team should realize that the direct end-to-end model may perform worse than the staged system because the labeled dataset is so small and the model does not start with useful domain knowledge. In this situation, the safer choice is to encode expert insight in the intermediate components, since much of the system's knowledge may need to come from that source.

Key points:

• Only a small number of labeled examples are available.
• The end-to-end model begins without domain-specific knowledge.
• It may not match the performance of the staged system.
• Expert understanding should be built into the component-based design.

Rubric: The response should identify the danger created by limited labeled data, compare the end-to-end model with the staged system without promising a certain result, and recommend relying more on expert knowledge encoded in component-level design.