Criterion: the AI interview coach that shows its work

Make the AI's judgment legible and accountable

An unexplained verdict teaches nothing. If a tool tells you that you scored a 6, you learn a digit. You do not learn what you said, why it counted, or how to do better next time. So the design problem was not "score an interview." It was "make a machine's judgment something a person can inspect and argue with."

Five components carry that argument. Each one below is the real, shipped component, rendered inline from the production markup, not a mockup. Each is paired with the decision behind it: what I tried, what broke, and why the shipped version won. The sample content is the app's own demo interview for a Design Engineer screen, so what you see is what a hiring manager sees if they land on the scorecard cold.

A score you can read at a glance, and one worth arguing with

The contrast with a bare "6.5" is the entire pitch. A score arrives with its state, its confidence, the weight it carries, and the reason it landed where it did. The most contestable score wears a coral frame, so the eye goes to the number most worth challenging.

Shows its work, made literal: the exact words behind the score

Open any score and it tells you the words that earned it. Each piece of evidence is a quote, the reason it moved the number, and a timestamp that replays the moment it was said. Direction is a glyph and a written note, never color on its own. Hover a quote and the same words light up in the transcript below, so the rationale and the source are unmistakably the same thing.

The signature move: contest a score, watch it re-evaluate in the open

This is the product's whole argument made operable. If you disagree, you say why, and the model re-evaluates against the current value. It can raise, hold, lower, or escalate to a human. The full contest record stays on the page. Nothing is rewritten in the dark.

A live interview that shows its read while you are still in it

The interview does not hide the machine until the end. You answer in text, and a running read updates on submit, scored against the same criteria the final card uses. It is labeled provisional on purpose, and it says so plainly: the read is reconfirmed at the end against the full transcript. The feed keeps the whole exchange in view, so nothing about how you are being read is happening off screen.

The interview replayed, wired to the scores

The scorecard answers "what did each criterion earn." The debrief answers the more useful question: "what did each thing I said drive." Every answer carries chips for the criteria it fed, marked as helped or hurt, and each chip jumps to that score above. It is the same evidence graph the scorecard renders, walked from the transcript side, so an answer can never claim to have driven a score the cards do not also show.

How I will know it is working

Criterion is live and early. I would rather state what I am watching for than borrow a number that does not exist yet. These are the signals that tell me the thesis holds in real use, not in a demo. This is the measurement plan, written before the data, because what you choose to measure is itself a design decision.

Sharpening one thesis, not stacking features

Every version answered the same question more honestly than the last: can a person read this judgment and argue with it. The work was subtraction as often as addition. Each step below maps to a component you just saw.