Case context: A startup has built a model that flags fraudulent transactions. Its evaluation shows the model still misses many examples even on the data used to fit it, so the main issue is underfitting rather than poor generalization. The team is considering two workstreams: Team A would increase model capacity, improve features, and tune optimization so the model can learn the fitted examples more accurately; Team B would focus on regularization, data expansion, and other methods intended to narrow the gap between fitted data and held-out data.

Question: Which workstream should be handled first, and why?

Sample answer: Team A should be handled first. When the biggest problem is underfitting, the most useful next step is to reduce the error on the examples the model is being learned from. Team B is designed to limit overfitting and strengthen results on new data, which is not the main problem described here. So the priority should be the changes that help the model match the fitted examples better.

Key points:

• Team A targets underfitting by helping the model fit the examples used for learning.
• Team B targets overfitting by improving behavior on held-out data.
• The first fix should match the model's current error pattern.

Rubric: The response should: 1. Say Team A should be prioritized. 2. Explain that Team A is meant to reduce underfitting and improve fit on the learned-from data. 3. Explain that Team B is meant to reduce overfitting by improving results on new data, so it is not the first priority here.