For 60 years, the dominant model in structured behavioural interviewing has been the standardised question bank: every candidate gets the same questions, scored against the same rubric, by the same panel. The standardisation is what produces predictive validity. The fairness, the bias-resistance, the legal defensibility — all of it rests on the consistency of the stimulus.
Standardised question banks made sense in a world without internet leaks, without Glassdoor question dumps, and without LLMs that can produce passable answers to any pre-published question in 30 seconds. They make less sense now. The question this piece tries to answer is: what replaces them, without losing the consistency property that made them work?
The standardisation paradox
Glassdoor’s interview-question database has roughly 4 million questions logged across 350,000 employers. The most-asked behavioural questions at major tech companies are well-documented enough that candidates rehearse specific “STAR-method” answers for specific question stems. Once a question is in the dump, it stops measuring spontaneous behaviour and starts measuring rehearsal quality.
LLMs amplify this dynamic substantially. A candidate can paste a known question stem into an agent, get back a polished STAR-method answer tailored to a plausible past project, and rehearse it. The rehearsal-vs-spontaneity signal collapses. The interviewer cannot easily distinguish a candidate who rehearsed an excellent answer from a candidate who is genuinely sharp.
The conventional response is to ask “surprise” questions — to make sure the candidate has not seen the specific question before. But the moment a hiring team adopts a new question, it goes into the next round of post-interview write-ups on Glassdoor, and within 90 days it is in the dump. Surprise as a design choice does not survive contact with the internet.
Construct-Templated Adaptive Interviewing
The design we have converged on at Basanite — and the one this whole piece is really about — separates the construct from the stimulus. The construct is the cognitive or behavioural quality you are measuring. The stimulus is the specific question you ask to elicit it.
Standardised question banks lock both. Every candidate gets the same construct and the same stimulus. The cost is the rehearsal problem.
Unstructured interviewing varies both. Different candidates get different constructs probed in different ways. The cost is the predictive-validity collapse that Schmidt and Hunter documented across 85 years of data — we covered the empirical work in our piece on structured interviews in the AI era.
Construct-Templated Adaptive Interviewing — CTAI, if you must abbreviate — locks the construct and varies the stimulus. Every candidate is probed across the same dimensions, scored against the same rubric, with the same evidentiary requirements. The specific questions asked are generated from that candidate’s own CV. Two candidates being assessed for the same role might be asked completely different questions on the surface; underneath, they are being measured on the same eight dimensions, against the same anchors.
Why CVs are the right anchor
A CV is a candidate’s privileged information. The candidate knows it; an LLM doesn’t. This asymmetry is the load-bearing property of the design.
Consider three questions. “Tell me about a time you disagreed with a senior engineer’s technical decision.” This is a standard behavioural question, well-rehearsed, and easily LLM-generated. “Tell me specifically about the migration from your monolith to your event-driven architecture at [Previous Employer], the one you wrote up in [their personal blog post] — what would you do differently?” This is the same construct (handling ambiguity and learning from experience), but the question depends on facts only this candidate knows. An LLM can hallucinate a plausible answer, but it cannot produce the candidate’s answer.
The CV-grounded question also produces richer signal even from honest candidates, because it forces them to engage with their actual lived experience rather than a generic narrative they have rehearsed. The answers tend to be more specific, more vulnerable, more diagnostic. Senior engineers we have run through CV-grounded round-one interviews routinely say things like “that’s the first interview question in years where I actually had to think.”
Fairness across very different candidate populations
One of the most important consequences of the CTAI approach has to do with fairness across candidate populations with structurally different CVs.
Consider two candidates applying for the same backend engineering role. One is an Oxbridge CS graduate with three years at a FAANG company. The other is a self-taught engineer who learned to code at 28, ran a successful indie SaaS for four years, and is now applying for their first traditional employment role.
A standardised question bank — “describe your most complex distributed system” — is structurally unfair to the second candidate. The Oxbridge graduate has been trained to answer that question; the self-taught candidate has built distributed systems but doesn’t have the vocabulary of academic distributed systems texts. The question doesn’t measure the construct (do they understand distributed systems trade-offs); it measures fluency in the vocabulary the question was written in.
A CV-grounded version of the same question — “tell me about the time your indie SaaS hit the database bottleneck and you migrated to a write-ahead-log architecture” for the second candidate, vs “tell me about the consensus debate inside your team when you migrated from a Raft to a Paxos implementation” for the first — measures the same construct in the candidate’s own vocabulary. The self-taught candidate has a fair shot at demonstrating depth. The Oxbridge candidate doesn’t get an unfair advantage from having internalised more vocabulary.
This is not a feel-good fairness argument. It is a predictive-validity argument: the variance you want your interview to expose is the variance in construct-level capability, not the variance in vocabulary. Standardised question banks contaminate the construct measurement with vocabulary measurement, and the contamination biases toward candidates whose background overlaps with the test-writer’s background. CTAI separates them.
How adaptivity works in practice
The adaptive element of CTAI is that the questions branch based on what the candidate has said. A candidate who claims deep expertise in distributed systems gets a deeper probe on a specific distributed-systems decision. A candidate who emphasises product judgment gets probed on a product trade-off they navigated. The interview goes where the candidate’s strongest claims are, then tests whether those claims hold up under specific probing.
This serves two purposes. First, it produces the deepest signal where it matters most — strong candidates have areas of demonstrated depth, and those are the areas where the interview should pressure-test. Second, it deters rehearsed answers, because the candidate cannot rehearse for a question stem they don’t know in advance.
The implementation is reasonably technical. The interview agent has the candidate’s CV in context, has the underlying dimension rubrics in context, and has a question-generation harness that produces specific questions whose form depends on the CV but whose function maps onto a specific dimension. A few hundred lines of prompt engineering and a careful set of rubric anchors are enough to make this work in production. We unpack the engineering details elsewhere; the conceptual point is what matters here.
What the candidate experience is like
From the candidate side, a CV-grounded interview feels less like an interview and more like a structured conversation with someone who has actually read their CV. The questions feel specific. The follow-ups feel relevant. The interview does not feel like a Glassdoor-rehearsed checklist.
Candidates who are good at the job tend to enjoy the experience. They get to talk about work they have actually done, with someone who is asking the right follow-ups. Candidates who rehearsed STAR-method answers for the standard question stems are less happy because their preparation doesn’t pay off as cleanly. This is the desired behaviour: the interview should reward candidates who have something real to say, not candidates who have rehearsed well for the wrong questions.
We collect candidate feedback on every Basanite interview. The most common comment is some variant of “that felt like the interviewer actually cared about the work I have done.” That is the design intent. The signal generation comes from the same place.
Where CV-grounded interviews don’t work
Two limitations to be honest about.
First, the design requires a CV that contains real substance. A candidate with a thin CV — graduate hire, career-changer with no relevant projects, junior with mostly course work — cannot be probed in the same way. The fallback for these candidates is a hybrid design: half CTAI on what they have done, half standardised behavioural on hypothetical scenarios. The signal is weaker for entry-level candidates because the input is weaker, which reflects the reality of evaluating entry-level candidates regardless of the interview design.
Second, the design depends on the CV being accurate. A candidate who has inflated their CV — claimed responsibility for projects they didn’t lead, exaggerated outcomes, fabricated technologies — will struggle to maintain a specific narrative under specific probing. This is sometimes useful (it surfaces CV inflation) and sometimes painful (a competent but self-effacing candidate underrepresented on paper can be probed in ways that don’t match their actual capability). The first case is desirable. The second case requires interviewers to probe down to the floor of the candidate’s self-described experience, not above it.
The relationship to anti-cheating
We have been writing about anti-cheating design across several pieces in this series — see our piece on what AI cheating actually looks like in 2026 and the practical playbook. CV-grounded interviews matter for the anti-cheating story because they remove one of the cheating vector’s key affordances: the ability of an LLM to produce a plausible answer to a question the LLM has never seen before. An LLM can produce a plausible answer to “tell me about a time you disagreed with a senior engineer” because it has been trained on countless variants of that question. It cannot produce the candidate’s answer to a question about a specific project that the candidate worked on and the model has no privileged knowledge of.
This is a structural property of CTAI rather than a stack of detection controls layered on top. The cheating-resistance is in the design, not the proctoring. That is the property you want.
What to take away
If you are designing a 2026 interview process, the CV-grounded principle is one of the cheapest, highest-leverage moves you can make. You don’t need to build it as elaborately as Basanite has — even a half-day exercise where a senior engineer skims each candidate’s CV before the interview and writes three CV-specific follow-up questions will substantially raise the signal floor of your round-one interview.
If you do want the full CTAI experience automated end-to-end — adaptive questions generated per candidate, scored against a consistent rubric, with quote-grounded evidence for every score above the rubric’s neutral midpoint — that is roughly what Basanite is. The FAQ explains the mechanics. The AI-resistant interview design piece covers the complementary round-two design. The structured interview piece goes deeper on the empirical literature underneath this whole approach.