How the accuracy check works in Interview Signal

The accuracy check is the AI model's reading of whether a candidate's technical claims hold up. On technical, coding and system design questions, Interview Signal picks out up to six claims from the latest answer, confirms each quoted phrase is really in what the candidate said, and marks it in the transcript: green if it held up, red if it did not, with a one-line reason when you rest the pointer on it. A strip above the follow-ups sums up the answer. It is a reading for you to check, never a score, and never in the submittal. The toggle is Check technical answers, under AI. This describes version 0.9.10, current as of October 2026.

When does the accuracy check run?

It runs with each round of follow-up suggestions, in the same request, so it adds no wait. Whether it runs depends on the question on the table:

• If the pinned question's category is Technical, Coding or System design, it runs.
• If nothing is pinned, it runs and the model decides from the answer itself, returning nothing for an answer with no factual claims, such as a story about a team, a preference or logistics.
• If the pinned question is Behavioural, Experience, Motivation, HR or Logistics, or one of the "Follow-ups you took", it does not run, because those answers rarely have a right or wrong to judge.

What does it show?

A strip above the follow-ups shows a verdict on the latest answer: Holds up, Partly right or Does not hold up, beside the caption "the model's reading of the technical claims — check it against what you know". Below it, each claim is listed with a ✓ or ✗, the candidate's own words in quotation marks, and a reason of under twelve words. The × hides the strip.

In the transcript, the same phrases are highlighted where they were said: green held up, red did not. Rest the pointer on a mark to read the reason. Marks from earlier answers stay as the interview goes on, up to 60 per interview.

How does it avoid marking things nobody said?

The model has to quote each claim word for word, 3 to 12 words, from the conversation. Before anything is shown, the app checks every quote against the candidate's own words from the last two minutes, using the same rule the scorecard uses for evidence. A quote that cannot be found is dropped. If no quote survives, nothing is shown at all, not even the verdict.

The model is also told to be conservative, because an unsupported red mark in front of a recruiter can cost a good candidate the role. A vague or unverifiable claim is not marked either way, and it judges at most six claims per answer.

How should I use a red mark?

1. Rest the pointer on the mark and read the model's reason.
2. Ask the candidate to walk you through it, for example: "Can you take me through how that worked in practice?" A real misunderstanding usually shows; so does a slip of the tongue.
3. If you want a second reading, type the claim into Ask the AI, such as "Is it true that a primary key can be null?"
4. Decide yourself. You, or the hiring manager you are screening for, know the field; the mark is a prompt to check, not a verdict.

A green mark is not proof of competence either. It says a stated fact was right, not that the candidate can do the job. Our guide to interviewing for technical roles as a non-technical recruiter covers judging evidence quality without judging every fact.

Does it affect the score or the submittal?

No. The accuracy check changes no score and is never written into the submittal or the scorecard. Scoring is a separate step that rates answers against your rubric; see the scorecard with verified quotes. The marks are saved with the interview in its folder on your computer, like the transcript, and are not among the counts sent to Interview Signal.

How do I turn it on or off?

Open AI in the live screen's top bar and tick or untick Check technical answers. Its description in the menu reads: "On technical, coding and design questions, the model marks the claims it can verify in the transcript — green held up, red did not — with a one-line reason. Its reading, for you to check; never a score, never in the submittal." It is on by default and the choice is remembered. Turning it off also hides the marks already in the transcript.

With Dual AI on, only the first model judges claims. Two readings of the same claim on one screen would be a debate, not a check.

Limits

• It is a fast model's view of a specialist fact, made mid-call. It can be wrong, especially on niche or very recent technology.
• It depends on the transcript. If technical terms were misheard, the quote may not match, and nothing is marked.
• It judges only the latest answer in each round, and only claims stated as facts.
• It needs a connection and minutes left on a metered plan, like the other AI features (see pricing).