The questions buyers ask.

No. It replaces the first technical round. The rounds after it stay yours, and they get better: the candidates who reach them have already accounted for their own code.

Engineering judgment: how a candidate diagnoses a problem in code they did not write, the decisions they make, how they check their work, and whether they can account for it. Not whether they can produce code — AI does that now.

A coding test checks the answer, and AI can produce the answer. Sphinque checks the answer, the record of how it was reached and the candidate's own explanation of it — and whether the three agree.

The interview is about work the candidate finished an hour earlier, in a codebase built for your role. Its questions come from their diff, their history and their AI transcript. The interviewer never knows whether their code passed, and the recommendation comes from a fixed rule over quoted answers, not from a model's score.

Backend roles on Java (Spring with JPA, Spring Data JDBC, Spring with MyBatis) and Python (Django, FastAPI, Flask), and mid-level frontend roles on React with TypeScript. The stack is read from your job description. Other languages and full-stack roles are not generated today.

Not today. Assessments are built from your job description, and you approve each one before it exists.

Not today. You invite candidates and read reports in Sphinque. If writing results back to your ATS matters, tell us on the walkthrough.

They read two tickets and the code and write a short plan. Then they fix a defect and build a missing feature in VS Code, on their own isolated machine. Nothing in the codebase says it is an assessment.

About 87 minutes: 7 to plan, 60 to build, 20 for the interview. The build time is set per role.

Yes. Claude Code works in their editor and terminal, and asks what they want changed before it writes code. A chat panel answers questions about the code. Every prompt, edit and command is logged as part of the evidence.

Everyone invited to an assessment gets the same codebase, byte for byte. That is what makes them comparable. The interview differs, because it is built from each candidate's own work.

It never counts against the candidate. A run that cannot show its tests executed halts instead of grading anyone.

Everyone who submits. There is no score gate — “walk me through what happened” is signal even when the code is wrong.

The candidate's own session: what they changed, why, how they checked it, what they asked the AI. No trivia, no whiteboard algorithms.

No, by design. An interviewer who knows the answer steers the candidate toward it.

A model judges what was said about each piece of work, and every judgment must quote the transcript. A fixed rule turns those judgments into the recommendation — the same rule for every candidate. No audio or video reaches it.

Help is allowed and logged, so that is the wrong question. The interview shows whether the candidate can account for the work. Where the code, the record and their explanation disagree, the report shows the disagreement and the quote.

A browser, a microphone and a camera. The interview is recorded so your team can watch it.

The recommendation and the reason for it, the hidden-test results, what the interview verified and the sentence each finding rests on, what it did not reach, how the candidate used AI, and a replay of the session. Your team records its decision next to it.

Test results are facts. Interview findings are judgments, and each one quotes its evidence. Anything the interview did not reach is marked as not reached and never reads as verified. Where a model contributed to a number, the report labels it.

Evaluation starts when the candidate submits. The report is complete once the interview has been analysed.

The interview is — screen share, camera and microphone, as one video your team can watch. The coding is not filmed; we keep the work instead. Activity telemetry is captured too: tab switches, focus, pastes, IP address and browser. All of it is stated to the candidate first, and no recording reaches scoring.

A versioned notice of exactly what is kept and recorded, which they acknowledge before their machine starts and again before the interview.

No. A single-use invitation link, valid for seven days by default, carries them through.

The hiring company. Sphinque processes it on your instructions. The privacy policy has the details.

No.

Not today. The security page says so in writing and lists the controls that do exist, with the values they run on.

On its own isolated machine, one per session, with the network closed except one pinned route to our model proxy. The machine is destroyed at the end. Hidden tests run afterwards on a separate machine.

Per active role and per verified screen — a candidate taken through the whole screen. Plans are on the pricing page.

Yes. Start with a pilot on one role, run alongside your current first round.

No. Plans are month to month. Annual billing is available by arrangement.