Skip to content
PROOFOS
Back to home

How it works

Know exactly what is measured before you agree to it.

Most tests never tell you. Here is the whole thing in four steps, plus what we look at and what we refuse to look at.
  1. 1

    We build a real task

    From the job advert, or from the kind of work you picked. Real data, working tools, and four mistakes hidden inside it on purpose.

  2. 2

    You work, with an AI teammate

    It looks things up and gives you confident answers. Four of them are wrong. You can see which tools it opened, so the clues are always there.

  3. 3

    We write down what happened

    Not opinions. Specific moments, quoted in your own words, each stamped with a time and a fingerprint so it cannot be changed later. The camera adds only counts, never footage.

  4. 4

    Scores are calculated from that

    Six skills, worked out from the moments we wrote down. Nothing is stored as a score. Ask why you got a number and you get the list.

What we look at

  • The work you wrote, and every message to your AI teammate
  • Which of the AI's mistakes you spotted, and which you did not
  • How often you asked it for evidence before believing it
  • Your answers in the trust quiz, and how confident you were
  • Whether you can explain your own choices out loud
  • Counts: pastes, edits, time taken, times you left the test
  • Camera counts, worked out on your device: seconds the picture was blank, no face in view, or eyes off the screen

What we never look at

  • Your face as an identity, your voice, or your screen. Camera frames never leave your device.
  • How you sound, your accent, or how fluent you are
  • Your background, your education, or where you live
  • Your mood, your personality, or anything guessed about you
  • Anything at all outside this one task

Which parts the AI is not allowed to decide

An AI builds the task and points at things you said. It does not set your scores. Anything you could reasonably argue with is calculated, so you can check it and so it comes out the same every time.

Did the AI's mistake end up in your work?
Decided by comparing text, letter by letter. No AI involved.
How did you do on the trust quiz?
Plain arithmetic against the real answers, which never leave our server.
What are your six skill scores?
Calculated from the recorded moments, by a formula anyone can check.
Should this person be hired?
We never answer this. There is no code anywhere that decides it.
Technical details, for developers and reviewers

Built on the Google Gemini API. Each part of the system uses the capability that actually fits it rather than sending everything to one model. Every instruction we send lives in a single file, lib/prompts.ts, so a disputed result can be argued about against something readable.

WhereModelCapability
Task design and markinggemini-3.8-flashStructured output, fixed seed, high reasoning
The AI teammategemini-3.8-flashTool calling, then streaming
Reading job advertsgemini-3.8-flashPDF and image understanding
Live statisticsgemini-3.8-flashWeb search with citations
Spoken answersgemini-3.5-transcribeSpeech to text, audio then deleted
Reading questions aloudgemini-3.1-flash-tts-previewText to speech
Matching skills to jobsgemini-embedding-2Embeddings
Live call logrefreshing

Nothing yet. Use the product in another tab and come back.