How the AI Interviewer Scores Your Answer
Behind the scoring: the rubric, the finite state machine, and how each answer becomes a number and a hiring recommendation.
Every answer you give in an AI technical interview is scored the same way a human interviewer scores it — against a rubric, turn by turn — and a finite state machine decides what happens next. Here's the actual mechanics, category by category.
The per-turn rubric: 0 to 10
Each of your turns is evaluated on three sub-scores that combine into a 0–10 score:
- Correctness & accuracy (0–4). Is the answer right? Do the facts hold, is the logic sound, does the code do what it claims?
- Depth of understanding / reasoning (0–3). Beyond right — do you know why? The mechanism, the trade-off, the complexity bound. This is where "answers" become "understanding."
- Communication clarity & structure (0–3). Could the interviewer follow it? Did you lead with the answer, structure the explanation, and speak in plain terms?
A strong turn scores in the 8–9 range; a correct-but-thin one lands mid-range; a confident-but-wrong one scores low regardless of how it sounds.
The finite state machine decides the flow
The interview is a deterministic state machine, not a free-for-all conversation. It moves through stages in a fixed order:
INTRO → WARMUP_QUESTION → CORE_QUESTIONS → CODING_CHALLENGE → FOLLOW_UPS → WRAP_UP → REPORT_GENERATED
The number of core questions depends on difficulty (3 for junior, 4 for mid, 5 for senior), and coding challenges similarly (1 junior/mid, 2 senior). The machine caps follow-ups at two per question — so a question never spirals forever, and the interview stays structured no matter how the conversation drifts.
Here's the part that makes it adaptive: after the coding challenge, if your average score is below 7.0 and you had a core-answer at or below 5, the FSM inserts a FOLLOW_UPS stage to probe the weak area before wrapping up. Score well, and you get a shorter, sharper interview; struggle somewhere, and the interview goes deeper on it. That branching is what makes the evaluation feel human.
The action loop
After each answer, the system chooses one of four next moves:
- advance — you answered solidly; move to the next topic
- follow_up — probe deeper, clarify, or challenge
- escalate_difficulty — you're doing well; raise the bar
- wrap_up — the interview is done
Each turn also records a follow-up type — clarify, probe deeper, challenge, redirect, affirm-and-advance, or connect — which is how the interview re-engages rather than just marching down a list.
The end-of-interview scorecard
When the interview finishes, every scored turn is collected and averaged to a 0–10 score, then normalized to a 0–100 overall score. The full transcript — your answers, per-turn scores, evaluator notes, and code — is synthesized into a report with:
- Five radar metrics: Correctness, Depth & Internals, Communication, Problem Solving, and Code Quality
- Strengths and growth areas, derived from the transcript
- Next study topics, specific to where you lost points
- A recommendation mapped from the overall score: STRONG_HIRE (85+), HIRE (70+), LEAN_HIRE (55+), LEAN_NO_HIRE (45+), NO_HIRE (below)
The radar isn't decorative — it's the same shape a hiring manager builds from a real loop, and it shows you where you lost points, which is the part that actually helps you improve.
Why the evaluation is honest
Everything in the scorecard traces to the transcript: a strength is a strength because a specific turn scored high, a growth area exists because a specific answer fell short. Nothing is invented — the numbers come from the per-turn rubric and the FSM's deterministic transitions. That's what makes the feedback usable: you can see the exact turn that cost you the depth score, and go study it.
The interview answer
"The scorecard is built from per-turn scores — correctness, depth, and communication, combined to 0–10 — averaged and normalized to 0–100. The interview itself is a finite state machine with deterministic transitions and difficulty-based question counts, capped at two follow-ups per question. If my average dropped below 7 with a weak core answer, the machine adds a follow-up stage to probe it. The final radar — correctness, depth, communication, problem-solving, code quality — maps to a hiring recommendation at the 85/70/55/45 thresholds."