Direct answer
What you can rely on: every sentence a report presents as something you said really appears in your answer transcript — each quote is checked, and if one doesn't match, the whole report is thrown away and generated again.
What you have to judge yourself: the transcript can contain speech-recognition errors, and the diagnosis, the better answer, the six ratings and the verdict are all model judgements — not independently verified, and not any employer's signal.
So use it this way: listen to the recording, read your own words, and treat the ratings as pointers to where to look, not as a verdict on you.
Editor's note: the author of this article is the author of the product (first-party content). The behaviour described below corresponds to src/lib/scoring.ts (validateScore, the quote check) and src/lib/report-service.ts (evidencePlayback, locating the recording).
1. How the rule is enforced
| Step | What happens | When it can't be done |
|---|---|---|
| Quote check | Every evidence quote, and the “your words” of every per-question note, must be found in the answer transcript. Before comparing, only extra whitespace is collapsed and letter case ignored; punctuation and wording must match. | If even one quote is missing, the scorecard is discarded and generation is retried once with the reason. If it fails again, the report shows as failed with a retry button — unchecked quotes are never shown to you. |
| Locating the recording | The quote is looked up in your recorded answers, and its playback position is estimated from where it sits in that answer's text. | If no recording contains it, the quote still appears in the report, just without a player. |
| Model answers | If you viewed a model answer for a question, the scorer is told, and that question's note says so. | — |
2. Four things it cannot guarantee
- The transcript is not your voice. Speech recognition occasionally mishears a word, so a quote being in the transcript doesn't prove you said exactly that — which is why evidence comes with the recording to check. Our own published test kept several mishearings exactly as transcribed.
- A judgement is not a fact. The diagnosis, the better answer, the six ratings and the three-tier verdict are model-generated practice feedback; they are not independently verified and are not any employer's hiring decision.
- Playback positions are estimates, calculated from the quote's share of the answer's text rather than per-sentence timestamps, so you may need to scrub back or forward.
- It isn't a lie detector. One experiment found that follow-up questioning actually increased the degree of faking (Levashina & Campion, 2007). The report helps you find your own weak spots; it doesn't judge whether your experience is true.
3. Why this way of practising is still worth it
One of the most-cited meta-analyses in personnel selection found that structured interviews — fixed questions and a shared scoring standard — predict job performance markedly better than free-flowing conversation (Schmidt & Hunter, 1998); a more conservative re-estimate in 2022 found the same direction (Sackett et al., 2022). That supports practising dimension by dimension, against evidence. It does not show that any particular AI report's ratings are accurate.
- Schmidt & Hunter (1998), Psychological Bulletin — DOI
- Sackett et al. (2022), Journal of Applied Psychology — DOI
- Levashina & Campion (2007), Journal of Applied Psychology — DOI
4. Check a report yourself in three steps
- Pick a criticism you disagree with, play its recording, and check that what you said matches the quoted sentence.
- If the recording and the transcript differ, trust the recording and discount that criticism. If they match, go to the question-by-question note and rewrite the better answer with your real experience.
- Don't rely on one report. Run another session with the same interviewer — the report compares it with your last attempt, and a weakness that keeps coming back is the one worth fixing.