Criterion — the AI interview coach that shows its work

AI assessment usually arrives as a black box — an opaque number with no traceability and no recourse.

For something as high-stakes and personal as interview prep, a bare score does the opposite of help. "6.2 — Communication" tells you nothing you can act on. You can't see which sentence cost you the point, you can't tell whether the system even heard the strongest thing you said, and you certainly can't argue with it. Worse, an unexplained verdict quietly trains the wrong lesson: people start performing for an invisible grader instead of getting better at the actual thing. The first design decision was to treat opacity itself as the bug to fix.

"Challenge score" isn't a feedback gimmick. It's the interaction that turns a graded subject into a participant in the evaluation.

When you challenge a score, Criterion re-reads your transcript against the same rubric, in the open: "Re-evaluating against the transcript & rubric…", then it tells you what it found. Crucially, the outcomes are honest — a challenge can revise the score up, hold it where it is, lower it, or return human review recommended. Roughly one in four challenges hold or lower the score. That asymmetry is the whole point: if challenging only ever flattered you, it would be theatre. Because it can also disagree with you, a revision actually means something.

Video exists to practise presence and nerves. You are scored only on what you say — never how you look, sound, your accent, or your background.

A real interview is partly a performance under pressure, so the prototype puts you on camera with a live interviewer to practise that. But the moment a product scores a face, the glass-box thesis is dead — and so is the trust. The load-bearing decision was drawing a hard line: the camera is for rehearsal, the rubric only ever reads the transcript. The concrete expression of that boundary is the delivery-notes panel — opt-in coaching on pace and filler words, framed in indigo, explicitly never scored, never written into the transcript, and never shown to a human reviewer. The boundary is stated in the UI, not buried in a policy page.

Each version defends the same idea — make the AI's judgment legible and contestable — against the easier, less honest version of every feature.

The arc isn't a changelog. Early versions found the product identity and put the challenge interaction at the centre. The middle versions made the interview a real, practised experience and added the trust layer that keeps video honest. The late work is where the thesis got defended: honest defaults, a replayable challenge, a warm dark theme that keeps "Criterion at night," and accessibility done as real contrast math rather than a checkbox.