How it works. A structured interview, run the same way every time.
Live is not a form and it is not a coding test with a webcam attached. It is a conversation that follows the structure the Interviewer’s Guide recommends, with the workspace the interview type actually requires, scored on a rubric you can read.
You configure the job
Three decisions, and one thing that is fixed.
The role
Title and requisition. Placing a role in an org hierarchy, so that results roll up, is planned and not built yet.
The level
Junior/Mid, Senior, or Lead/Staff+. This sets the weights the score is computed with, the interviewer’s prompting budget, and the bar it holds. A junior who needs a nudge on non-functional requirements is performing at level, not failing.
The interviews
Screen, Coding, System Design, Data Science, or any combination, at 45 or 60 minutes (25 or 30 for the Screen), optionally gated so the later ones only release if the screen clears a threshold. On Coding you set the allowed languages and, optionally, a preferred choice of language that the candidate is free to ignore.
The interviewer
Always Bracket, on every plan. Choosing a manner or a voice is planned and not built yet; when it ships it will change how Bracket sounds and never the rubric, the difficulty, or the score.
The candidate gets a link, not a chore
No account, no password, no download, no scheduling back and forth. One click opens a lobby where they check their devices, read exactly what will happen and what is assessed, and give or refuse consent, item by item, with nothing pre-checked.
Nothing is recorded until they start. They can take a 60 second warm-up first so the first thing they hear is not the first thing being scored.
They can also decline and ask for a human interview instead, without penalty. That path is visible, not buried, because a consent flow with no real refusal is not consent.
What the candidate is told, before anything starts
- The interview is conducted by an AI, and it is named.
- Exactly which four things are assessed.
- That their face, expressions, tone of voice, and appearance are not analyzed.
- That their name is hidden from the scoring system.
- That a person at the company makes the decision.
- What is kept about them, for how long, and how to have it deleted.
- That assistance signals are planned and not built yet. When they ship, candidates will be told.
- Any notice their jurisdiction requires, injected automatically.
The interview follows a structure, and the candidate can see it
Most interviews run out of time before they reach technical depth, which means they measure requirements gathering and nothing else. Live puts the structure on screen and checkpoints every transition out loud.
Two numbers per phase: the 45 minute session, then the 60 minute one. Notice that the extra 15 minutes is not spread evenly. Requirements gains 2 minutes and the deep dive absorbs 8 or 9, because most interviews already spend too long on scoping and a session still in requirements at minute 25 has gone off the rails. Depth is what the extra time is for.
It waits
Bracket is bound to one facilitation rule: ask, then wait a full 30 seconds of genuine silence, then probe with a targeted follow-up rather than steering toward an expected answer. Silence tolerance can be tuned but never below 20 seconds.
It covers the rubric
The system tracks which dimensions have real evidence. When time gets short on an uncovered dimension, it forces the transition. An interview that never reaches depth is reported as Insufficient Signal, not as a low score.
Hints are an interviewer’s tool
The interviewer can offer a hint, and a candidate can ask for one. Hints are not stored as a separate record and are not shown in the report. The scorer reads the transcript, so a candidate carried to a good answer is judged on what was actually said.
Scoring, and how a number gets assigned
Four dimensions, weighted for the level you selected. Here are the written anchors for Architectural Judgment, which is the dimension most often reduced to gut feel. The scorer does not receive the anchors yet; scoring against them is planned.
| Level | Score 1 | Score 3 | Score 5 |
|---|---|---|---|
| Junior / Mid L3 to L4 | Names technologies with no reasoning. Cannot articulate an alternative. | Makes reasonable choices and explains the reasoning when asked. Recognizes at least one trade-off per major decision with prompting. | Volunteers trade-offs for major choices. Compares a real alternative. Reasoning is coherent even where the choice is conventional. |
| Senior L5 to L6 | Pattern-matches to a familiar architecture. Trade-offs named only after being asked twice. | Names the trade-off being accepted for most significant choices. Sound, but occasionally generic rather than tied to the stated constraints. | Names what is being accepted for every significant choice and ties it to a requirement. Revisits decisions when constraints change. Distinguishes reversible from irreversible. |
| Lead / Staff+ L7 and above | Reasoning is technology-first rather than constraint-first. No sense of cost, operability, or organizational fit. | Strong trade-off reasoning inside the system boundary. Some blind spots on cost, migration path, or team capability. | Reasons from constraints to architecture. Weighs cost, operability, migration, and organizational fit alongside technical fit. Identifies the one-way doors. |
Every quote is checked
The scorer cites quoted moments from the transcript, and each quote is checked word for word against it. A quote that does not match is discarded. When too few verified quotes are left, the report says insufficient signal instead of giving a score.
One scoring pass, for now
Scoring is one model call, and the confidence shown is the model’s own figure. Evaluating each session several times and taking the spread as the confidence is planned, and not built yet.
The total is arithmetic
The model produces four dimension scores. The weighted rollup happens in code. A model that can be argued into a different total is not a rubric.
You spend four minutes, not forty-five
The report opens on the recommendation and the four dimension scores. Each score carries a written rationale and the quotes it rests on. Click a quote and the timeline jumps to that moment in the transcript.
Below that: the pseudonymized transcript itself, with the exact moment behind each cited quote on the same timeline. The interview is recorded, audio and video, with the candidate’s consent from the lobby, and the recording is kept with the transcript for the retention period your organization sets. The scorer is given the transcript as text and, for System Design, the final whiteboard as text. It never receives the audio or the video.
Disagree with a score? Recording an override with a written rationale, kept beside the machine score and never replacing it, is planned and not built yet. Overrides are not stored today, and reports cannot yet be marked as reviewed.
Recommendation bands
The planned assistance signal will never change the band. Mixing them would turn an unfalsifiable suspicion into a score.
The fastest way to evaluate this is to run it.
Send one interview to yourself, then read your own scorecard with the evidence links.