You answer a theory question fast and correctly, and the interviewer's face does not change. The next question is short — "so what did that cost you?", "how much too high?", "show me where" — and you have nothing. The round goes quiet, the questions get easier rather than harder, and the interviewer moves to "do you have any questions for us?" ten minutes early. Afterwards the feedback says "theoretical" or "textbook knowledge, no depth", and you cannot see what was wrong with an answer that was factually right.
Any theory or viva round, most often the second technical interview. It bites hardest on the metrics, leakage and overfitting questions, because those have tidy one-line definitions that are easy to memorise the night before. Students who prepared from a question list are more exposed than students who prepared from their own project, and the collapse always happens on the first follow-up, never on the first question. Linked module: 6.4 (ML theory viva bank).
A definition and an answer are different objects, and the interviewer asked for the second one.
"Why did you choose recall?" is a question about a decision you made. "Recall is true positives over true positives plus false negatives" is a fact about a formula. Both sentences are true, and only one of them is a reply. The interviewer is not checking whether the fact is in your head; she is checking whether the decision was in your hands.
The failure is invisible on question one, because a memorised line is fluent and sounds prepared. It becomes visible on the follow-up, which is why interviewers always ask one. Follow-ups almost never ask for another fact. They ask for a consequence ("what did that cost?"), a size ("how much?"), a location ("where in your code?") or a boundary ("when would that be wrong?"). None of those can be recalled. All of them can be answered from a run you actually did.
There is a second, quieter symptom that confirms the diagnosis: the numbers you cannot supply are your own. Not knowing the exact formula for PR-AUC is normal for a fresher. Not knowing what recall your own model reached is a statement about whether the project happened.
The reason this happens to hard-working students is that memorising feels like studying. Twenty definitions can be read in an hour, and twenty defended answers take a week. The hour feels productive right up to the moment somebody asks a second question.
Replace every memorised line with a four-part skeleton, and refuse to keep any answer that fails part three.
The skeleton, in order
The own-example rule. Write out your answers and mark every one whose part three has no number, no file and no decision in it. That list is not a weak-answer list; it is a list of things you did not actually do. Fix it by running the experiment, not by rewriting the sentence. If you cannot run it, remove the topic from your resume — see WR-159.
Rehearse the follow-up, not the answer. For every question, write the three follow-ups an interviewer would ask: what did it cost, how much, and when would this be wrong. Answer those out loud. If the answer to a follow-up is "I would have to check the notebook", open the notebook now and put the number in your head, because it is your notebook.
Prepare from your own work, not from a question list: your recap and check blocks in each module are already written in the four shapes an interviewer uses. Keep a one-page numbers card for every project — the metric, the baseline it beat, the confusion-matrix counts, the threshold you chose and why — and revise the card, not the definitions. Practise in pairs, because sitting in the interviewer's chair teaches you which answers have nothing under them faster than any amount of solo revision. Use a language model to interrogate you, never to write your answers; a generated answer fails on exactly the same follow-up a memorised one does. Before any round, run one test on yourself: pick any claim in your resume and try to defend it for ten minutes without opening a file.
Q: "You said your model used recall. Why?" Strong answer: give the decision in the first sentence, price both mistakes in rupees, quote your own measured numbers on both sides of the trade, and name the assumption you had to guess. Weak answer: the formula. The tell an interviewer listens for is whether the numbers in part three could belong to anyone else's project — if they could, they probably do.