The unit tests are green and the model is wrong. What does an ML test stack contain?
Answer it out loud before you open anything. The value of the flags below is in comparing them to what you actually said — including whether you asked about the data before naming a model.
A team ships a retrained model through CI with full test coverage on the training and serving code. Two weeks later they discover that a schema change upstream renamed a column, the feature pipeline silently filled it with zeros, the model trained on zeros, and the serving path served zeros. Every test passed and the offline metric dropped only slightly, within the range attributed to noise.
React to this
Say what you would question, what you would trust, and what you would need to know first.
Post-incident timeline (illustrative) day 0 upstream renames txn_amount -> transaction_amount day 0 feature job: unknown column -> default 0.0 (no warning, coalesce in SQL) day 1 CI: 412 tests passed; retrain job succeeded; validation AUC 0.862 (prev 0.868) day 1 promotion: "delta within noise" (no noise estimate on record) day 14 fraud analyst: "the model stopped catching high-value fraud" finding: txn_amount importance went from 1st to absent; all-zero column
What it is really testing
Whether the candidate knows that code tests test code, and that a model needs tests for its data, its features, its training behaviour, its invariants, its serving contract, its robustness and its regression against the previous version — and can say which of those would have caught this incident and at which stage.