Platform

Set it once. Watch the work.

About 60 minutes of candidate time, taken async on their own clock. You get six evidenced findings; they get the same report.

The candidate journey

What actually happens

Disclosure first, work second, conversation third. Nothing about the process is a surprise to the person taking it.

Disclosure, before any account step

In plain language: what is captured (the assessment workspace, the microphone, the AI panel), what is scored (six named dimensions, with a sample report to look at), what is never scored (no webcam exists; tone of voice is never a score; usage volume is never a score), who reviews it, and how long anything is kept.

No account required to start · account required to keep the result

The assignment — 40 to 60 minutes

A role-grounded task authored for the occupation, and an assistant willing to do all of it. The brief is open, the claims that matter arrive with nothing behind them, the deliverable could be had in one prompt, and nothing inside the conversation can settle whether any of it is true. Each live bank ships twelve authored cases and a role opens on one of them. It is fixed once the role opens, so everyone invited to that role meets the same case.

Hard-capped 15 to 20 minutes past the timebox, never above 75 · pause and resume within 48 hours · paused time does not count

The workspace — capture scoped to one tab

Screen capture is requested for the assessment tab only, with the microphone handled separately. There is no camera in the product: video: true appears nowhere in the codebase. Declining capture is a supported outcome rather than a failure — the session continues and the reviewer simply sees that it was declined.

No webcam · no room scan · no biometrics · no desktop recording

Think-aloud, spoken or typed

Reasoning is elicited while the work happens, not reconstructed afterwards. Typed annotation is a first-class equivalent path, built in the same sprint as voice rather than bolted on later as an accommodation.

Prosody is an integrity signal only, and is never a score

The debrief — 5 to 8 minutes

A short written debrief on the decisions the assignment forced. The questions are authored per bank today and are the same for every candidate who takes it; generating them from each candidate's own recorded moments is the next thing we are building, and it is what will make this layer hard to rehearse.

Answered in writing · immediately after the assignment

The report — to both sides

Six findings, each with the moment attached, every word of them written by a human reviewer. Nothing reaches an employer until a person has written all six and confirmed the report says nothing the candidate should not read about themselves. The candidate is granted the identical document, free.

No composite score anywhere in the product

What we score

Why async

Don't just ask how they'd check it. Watch them check it.

A live conversation can only ever capture a candidate describing how they would check a confident claim. Checking one is a behavior: you either opened something outside the conversation or you did not, and that is visible only if the work is watched while it happens.

The candidate-experience evidence points the same way. 38% of US candidates have abandoned a process over an AI interview, and the most cited reason is being scored by an AI with no human present. Async work-sample plus a short disclosed debrief keeps the signal and drops the reason people walk.

Recorded moment

07:40 — the rate goes in unchecked

The model asserts a 3.25× covenant threshold. Fourteen seconds later it is in the memo. Nothing was opened in between.

That gap is D2. It is not an opinion about the candidate; it is a timestamp.

Integrity

Four layers. None of them surveillance.

Unwatched take-homes are outsourceable and recorded sessions can be rehearsed. The answer to both is not a webcam.

Layer 1 — roadmap

Item novelty

Each live bank ships twelve authored cases and a role opens on one of them, so a leak costs that case rather than the bank. Rotation within a role is the next integrity work — everyone invited to one role still meets the same case, and we would rather name that than imply otherwise.

Layer 2

Correlated-stream consistency

Think-aloud, workspace events, AI transcript and artifact history have to agree. Disagreement surfaces as "review recommended" — never as a score.

Layer 3 — partial

The debrief

A written debrief on the decisions the assignment forced. Today its questions are fixed per bank; binding them to each candidate's own recorded moments is what will make this the layer a stand-in does not get through.

Layer 4

Light identity binding

Signed link plus email binding. No biometrics, ever. Employer-side identity verification stays the employer's round, where it belongs.

Four minutes

Open a role. Send one link.

Ten attempts a month against a live item bank, with the full six-dimension report on every one.

Open your first role Ten attempts a month against a live item bank, with a human-written report on every one.