Time in the story: 13:00
ContactTalk to us

A flag is evidence, not a verdict.

In an interview, Crivello notes a few things that can matter. A person reads the notes and decides. No note changes a score, and no note rejects anyone.

Built and tested. Not live yet.

  • In testing

  • In testing

  • In testing

  • In testing

  • Optional

  • Camera and gaze

    Building next. Switched off today, so it is in no note.

Every note says what happened, and when.

Plain words, the minute it happened, and a link to that moment in the recording. Never a guess at why.

Some notes are not about the candidate at all. A dropped network is filed under "Not the candidate", so nobody reads it as a choice.

Every check is still "In testing", and says so. A check in testing never counts toward the review level.

Integrity, coding roundExample

Score3.0 / 4

Signals for context, not verdicts. They never change a score or move a candidate.

  1. The round starts at 13:32. Nothing noted yet.
  2. 13:36Left the interview window for 9 seconds.Candidate activityJump to 13:36In testing
  3. 13:44Pasted 412 characters. Close to an answer an AI gave for this problem, lines 12 to 40.Candidate activityJump to 13:44In testing
  4. 13:51Network dropped for 20 seconds.Not the candidateJump to 13:51
  5. 13:58300 characters appeared in 4 seconds.Candidate activityJump to 13:58In testing

Notes: 0Score changed: never

A person reads it and decides.

The founder's team and ours see the same notes. One of them clears it, or books a follow-up call where the candidate can explain.

  • It never rejects anyone on its own.
  • It never stops or blocks the interview.
  • It never scores emotion, stress or the face.
  • It never says what someone meant. Only what happened.
Reviewer viewExample

Not rated yet. Every check is in testing.

3 notes from the coding round, 1 network drop. Try it: you are the reviewer.

No decision yet. The notes wait for a person.

The candidate sees the same notes.

On their "Your data" page, in the same words the hiring team reads. They can tell us their side, and a person answers within 5 business days.

They can also ask for a re-check, or for a human interviewer. Before every round they are told what we check.

Your data: integrity recordExample

What the interview rooms noted, in the same words the hiring team sees. These are notes for context, not verdicts.

  1. Left the interview window for 9 seconds.Coding round, at 13:36 In testing
  2. Pasted 412 characters.Coding round, at 13:44 In testing
  3. 300 characters appeared in 4 seconds.Coding round, at 13:58 In testing

Wrong for 1 honest person in 100, at most.

A false flag is a warning on an honest session. A check may only warn after at least 300 honest test sessions in each kind of room show it warns 1 honest person in 100 or fewer.

We count with a strict bound, so a few more sessions buy room for one more warning, never a looser rule.

The gate, per kind of room
  1. 300 sessions0 warnings allowed
  2. 473 sessions1 warning allowed
  3. 628 sessions2 warnings allowed
  4. 773 sessions3 warnings allowed
  5. 913 sessions4 warnings allowed

Today: honest sessions run0

So every check is In testing.

Same bar for every group: each needs 50 sessions first. Indian English, Hinglish, transliteration keyboards, screen readers and dictation, built-in and phone webcams, low light, glasses, laptops older than 4 years.

Method first. Numbers after, even bad ones.

We publish how we measure before the first outside tester. Then a trust report on test sessions, labelled as test data, on the date we set, whatever it shows.

Every group is reported, including the ones with too little data. The gate above is public. The level at which each check notes something stays private, so nobody can tune to it.

Trust report: false flags by check
False-flag rates by check
CheckSessionsWarningsBound
No numbers yet. No check has passed the gate.
Publish plan
  • Fri 23 Oct 2026This page rebuilt. Every change to it is dated here.
  • Thu 5 Nov 2026Protocol v1: the rules, the sample sizes and the groups, dated before any outside tester.
  • Fri 29 Jan 2027Trust report, issue 1. Labelled "tester sessions, not live candidates".

Not done yet

A bias audit. We run an internal name, gender and language check while we build. It is not an audit, and we do not publish its results.

Ask how a check works.

Karan built them and will walk you through any note, any rule, any number. 15 minutes, no sales team.

Talk to Karan →Read security

Candidates: read the AI interview notice, or write to content@crivello.ai