Evaluation Drills

In one line: for each scenario, name the metric you would report and the trap you would call out before anyone asks.

Choosing the metric

1. A medical triage model flags patients needing urgent review. Which metric?

Recall, at a precision floor the clinical team can staff. A missed urgent case is unbounded harm; a false alarm is a review. Report precision at that fixed recall so it cannot be gamed. Trap: never quote either alone.

2. An email client auto-deletes suspected spam. Which metric?

Precision, very high. A deleted real email is unrecoverable from the user's point of view. Better still, do not auto-delete — a junk folder converts an irreversible error into a recoverable one, which is a design fix rather than a metric fix.

3. Your fraud model has 99.8% accuracy. Is it good?

Unknowable, and probably not. If fraud is 0.2% of transactions, predicting "never fraud" scores 99.8%. Ask for PR-AUC against its base rate, and precision at the recall the review team can absorb.

4. Two models: A has ROC-AUC 0.94, B has 0.91. Pick one.

Not yet. On a rare positive class ROC-AUC compresses exactly the differences that matter, because both its axes are normalised by class size. Ask for PR-AUC and for precision at the operating recall — B can easily be the better model at the point you actually run it.

Thresholds and calibration

5. The model is "not accurate enough". First move?

Sweep the threshold before retraining. Minutes rather than weeks, and it often recovers most of the gap. The threshold is not part of the model — it is chosen afterwards from the cost asymmetry.

6. A team rebalanced training data to handle a rare class, and downstream a policy multiplies the score by transaction amount. What breaks?

Calibration. Rebalancing inflates the output probabilities, so every expected-loss figure is too high and the policy over-blocks. Nothing errors and every model metric looks better. Fix: recalibrate against the true base rate on held-out data.

7. When can you ignore calibration entirely?

When only the ordering of the score is used — a ranked feed or search results. The test: is the score's magnitude read anywhere, or only its order?

Ranking

8. A search team has spent three months tuning the ranker and quality is still poor. What do you check?

Retrieval recall at the candidate-set size. If recall@1000 is 60%, final quality is capped at 60% and no ranking work can pass it. Measure the stages separately — recall for retrieval, NDCG for ranking.

9. Which ranking metric for a "find my order" search box, and why not NDCG?

MRR. There is exactly one right answer and everything after it is irrelevant, which is precisely MRR's assumption. NDCG's graded relevance buys nothing when relevance is binary, and it costs annotation.

10. Your recommender's offline NDCG improved 8% and the A/B test is flat. Explain.

Expected. Offline judgements exist only for items the old system showed, so the metric rewards agreement with it and cannot see wins from newly surfaced items. Use offline to filter candidates, then interleave, which detects a difference on far less traffic than a full experiment.

Generation and judges

11. A summarisation system scores well on word-overlap against references but users complain. Why?

Overlap metrics measure surface similarity, so they punish correct paraphrase and reward fluent text that reuses reference vocabulary while being wrong. Use them as regression detectors only. Measure faithfulness against the source instead.

12. You use the same model family to generate and to judge. What is wrong?

Self-preference bias, worth 10–25%. The system is grading itself and will report that it is doing well. Judge with a different family, and validate the judge against human labels on a sample.

13. Your pairwise judge says the new system wins 68% of the time. What do you check first?

Position bias. Judges can prefer the first-presented answer up to 75% of the time. Re-run with the order swapped: if the verdict flips, the comparison is noise. The flip rate is itself a useful measure of how much of your evaluation means anything.

Experiment design

14. The dashboard went significant on day three. Ship?

No. Peeking inflates the false-positive rate far above the nominal level, because every look is another chance for noise to cross the line. Either fix the duration in advance or use a sequential method built for continuous monitoring. Also check sample ratio and plot the effect over time — a declining effect is novelty, not a win.