AI in hiring should show its receipts.

A human makes every decision.

AI gathers evidence and drafts a rationale from it. It never writes a score, never ranks or compares candidates against each other, and never advances or rejects anyone — on any plan, in any mode. In the code, `scores` has exactly one writer, it stamps the human who set it, and the decision field is only reachable from a person clicking a button. On Expert and above an interviewer can switch on an advisory recommendation about a single candidate; it is described below, and it is still a person who decides.

Recommendations are advisory, and off until you turn them on.

Available from the Expert plan up, off by default, and switched on per interview rather than per workspace — so turning them on for one candidate does not turn them on for everyone. When on, the model writes one short note about what the evidence supports and which moments are worth re-reading. It is instructed never to give a hire or no-hire verdict, never to compare the candidate to anyone else, and never to score them, because that output is what would make this an automated decision system rather than a document a person reads. The note is labelled advisory wherever it appears, the toggle writes an audit event naming who flipped it, and the report records whether one was present when it was drafted. Depending on where you hire, automated recommendations carry disclosure and audit duties — NYC Local Law 144 and the EU AI Act are the usual two — and that is your obligation as the employer, not ours.

Every score links to evidence.

One tap from any number to the moment, message or piece of work that produced it. A score cannot be saved without it: it's enforced in the application, by a NOT NULL column, and by a database constraint that also rejects a blank.

Integrity signals are questions, never verdicts.

Paste, tab-switch and cadence detection prompt a conversation. They never label anyone a cheat — a signal is a reason to ask, not a verdict. No exercise hands the candidate an AI to answer for them — except one, Working with AI on the paid plans, where the assistant is the task and the candidate is marked on how they direct it. If you're testing judgement, a ghostwriter defeats the point.

Candidates are told how it works before they start.

Every candidate page we serve says an AI is involved, and each simulation names the counterpart as AI on screen. What we cannot control is what you say in your own invitations and calls — the Terms make that your obligation, and the invitation we draft for you includes it. A human reviews everything the model produces here; nothing is judged by machine alone.