AI models are learning to hide their reasoning. The question that grades yours.
On September 7, 2026, TechCrunch published a glossary of AI terms worth knowing, and one of them describes a change in how machines think. Opaque recurrence is the name for a model that "repeatedly loops the same query through its internal layers, instead of reasoning step-by-step in plain language." The technique works. The problem is what it costs: it "leaves far fewer readable traces," and those logs, TechCrunch notes, are "a key tool for catching misbehavior."
So the reasoning still happens. It just stops being legible. The answer arrives with its justification compressed into something nobody can read back.
That is worth sitting with, because it is also the most common way a strong candidate fails a senior interview. The thinking was real. The conclusion was correct. And none of it was visible, so none of it could be graded.
Why they ask it
On its face this is a question about outcomes. It is not. The interviewer already assumes you picked something workable, because you are in the room. What they cannot assume is that you would pick well again, in a situation they have not described, against constraints they have not met yet.
The only evidence available for that is the shape of your reasoning. Which options existed. What separated them. What you were optimizing for, and what you were willing to lose. A candidate who supplies the conclusion alone is asking the interviewer to take the process on faith — and the interviewer's entire job, for forty-five minutes, is to not do that.
There is a second reason, quieter and more important at senior levels. The alternatives you considered reveal the size of the space you were searching. Two options means you saw a fork. Five means you saw a landscape. Interviewers infer your ceiling from the options you rejected far more than from the one you kept.
The trap
The dominant failure is the answer that is all destination. "We evaluated a few approaches and went with Kafka, and it scaled cleanly." Every word may be true and the answer is nearly worthless — it is a result with the reasoning looped away inside, unreadable.
The mirror-image failure is the tour. Six options, each described at equal length, none of them ever eliminated. That reads as a candidate who cannot converge, which at senior levels is the more expensive flaw: the organization does not need someone who can enumerate, it needs someone who can close.
Both failures come from the same missing piece — the discriminator. Not the list of options, but the specific property that killed each one. "Postgres handled the volume but not the replay requirement, and replay was the whole reason we were rebuilding." One sentence, and the interviewer can now see you choose.
Applying STAR-T
Situation. Set the constraint before you set the options. Options are only interesting relative to what was binding — a deadline, a cost ceiling, a compliance requirement, a team that had to operate it at 3am. Name the binding thing first and the rest of the answer has somewhere to stand.
Task. Say what you were optimizing for, in one line, and say it before you describe a single option. "I needed something the on-call team could debug without me." That sentence is the grading rubric you are handing the interviewer, and handing it over deliberately is a senior move.
Action. Three or four options, each with its discriminator, ordered so the reasoning narrows. Say what you actually did to distinguish them: the spike you ran, the number you measured, the person whose objection changed your mind. Then name the finalist and the runner-up, and be precise about what separated them. If you cannot articulate why the runner-up lost, you did not run a comparison — you ran a preference.
Result. Give the outcome, and give it a horizon. "It's been in production fourteen months and we've added two consumers without touching the core." Durability is the evidence that the choice was structural rather than lucky.
Trade-off. Name what the winner cost you, because every real choice bought something at a price. "We accepted a heavier operational footprint — one more system to patch and page on — in exchange for replay we could not get any other way." A candidate who names the price is a candidate who knew they were paying it.
The follow-up that breaks weak answers
Expect "what would have changed your mind?" This is the question that separates a decision from a rationalization, and it is very difficult to answer convincingly after the fact.
A real answer names a threshold that existed at the time: "If the replay window had been under a week, the cheaper option wins and we do not rebuild at all — I checked that number first because it was the only one that could have stopped the project." That is a candidate describing a decision with edges. The unconvincing version — "I'm not sure, it was clearly the right call" — tells the interviewer the alternatives were never live, and that the tour of options was a story told backwards from the answer.
Keep the traces readable. The reasoning is the part being graded.
Score your answer against the director’s bar
Q: Tell me about a time when you chose one solution from many possible solutions.
Bank a decision where you can still name the runner-up and what killed it, and rehearse it in L8 Loop until the discriminators are the loudest thing in the answer. Try it free →