The customer's marketplace has network effects, so a user-level A/B test is biased. How do you measure the model's impact?

The screen is whether the candidate recognizes that SUTVA is broken before proposing a design, and can name the failure direction: shared inventory or pricing makes a pricing model look worse than it is, while a recommender that trains on contaminated logs can look better. The reserved follow-up is the analysis side, that switchback observations are autocorrelated so naive standard errors understate uncertainty, and you need block-level or time-clustered inference; knowing the design is half, knowing the analysis is the staff half.