> ## Documentation Index
> Fetch the complete documentation index at: https://docs.exterview.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# How Scoring Works

> The exact method behind every Exterview score — how answers become competency scores, how competencies become a round score, and what a score is not allowed to conclude.

Every Exterview score is produced by one published method. This page is that method, in full. You should be able to take any report, read the competency table printed on it, and arrive at the headline number yourself.

That is deliberate. A score you cannot recompute is a score you cannot defend — to a hiring manager, to an auditor, or to the candidate it was about.

## The shape of a score

A score is built in two levels, not one.

<Steps>
  <Step title="Answers become a competency score">
    Every answer is scored 0–10 against the sample answer for that question. All
    the answers inside one competency are **averaged**.
  </Step>

  <Step title="Competencies become a round score">
    Each competency's average is weighted by the percentage you set in **Setup →
    Scoring**, then combined into a single 0–100 number.
  </Step>
</Steps>

### Why two levels matter

Because the number of questions asked in each competency is not a decision you made — it is an artefact of how that particular round ran.

If questions were weighted individually, a panel that spent 18 of 20 questions on Technical would silently turn your 50% Technical weighting into 94%. Averaging inside the competency first means **your configured mix is the mix that gets applied**, whether a competency received two questions or twenty.

Question count still matters — but it affects **confidence**, not weight. See [What a thin round can conclude](#what-a-thin-round-can-conclude).

## Worked example

Your competency mix: **Technical 50% · Communication 30% · Culture 20%.**

The panel asks 18 technical questions (candidate averages 8/10) and 2 communication questions (averages 3/10). No culture questions come up.

| Step                     | Calculation               | Result |
| ------------------------ | ------------------------- | ------ |
| Technical average        | 8 ÷ 10                    | 0.80   |
| Communication average    | 3 ÷ 10                    | 0.30   |
| Culture                  | not assessed              | —      |
| Weight actually assessed | 50 + 30                   | 80     |
| Weighted total           | (0.80 × 50) + (0.30 × 30) | 49     |
| **Round score**          | 49 ÷ 80 × 100             | **61** |

Two things to notice.

**Culture is removed from the denominator, not scored as zero.** A candidate is never penalised for a competency nobody asked them about. This is also why a candidate who answers everything perfectly scores exactly 100, regardless of which competencies came up.

**Technical does not dominate.** It was asked nine times more often than Communication, but it still counts for 50% — because that is what you configured.

## What counts as evidence

Three states, kept strictly separate:

| State                   | What it means                            | Counts toward the score?         |
| ----------------------- | ---------------------------------------- | -------------------------------- |
| **Not asked**           | The question never reached the candidate | No — shown in coverage only      |
| **Asked, not scorable** | Asked, but no usable answer to grade     | No — removed from both sides     |
| **Asked and scored**    | Graded 0–10                              | **Yes — including a score of 0** |

A poor answer scores 0 and counts as 0. It is never treated as though the question was skipped — that would reward a bad answer by removing it from the calculation.

## Questions asked outside your configured list

Interviewers improvise, and that is intended — a panel brief gives the interviewer a goal for the round, not a script. Off-script questions are real evidence and they count.

Each one is classified against your competencies:

<AccordionGroup>
  <Accordion title="Mapped to one of your competencies">
    Scored exactly like a planned question. There is no penalty for improvising
    — an off-script technical question counts the same as a planned one testing
    the same thing.
  </Accordion>

  <Accordion title="Classified with low confidence">
    Held back from the score and flagged for you to confirm or reassign. We will
    not score a candidate on a guess about what a question was measuring,
    because that is not something we could explain to them afterwards.
  </Accordion>

  <Accordion title="Outside your competency matrix">
    Shown on the report with its score, but excluded from the headline. If the
    same theme recurs across rounds, Exterview will suggest adding it to your
    matrix — as a recommendation you approve, applying to future rounds only.
  </Accordion>

  <Accordion title="Touching a protected characteristic">
    Never scored, in either direction — including when the candidate answered
    well. The question is flagged for review. A protected characteristic cannot
    help or harm a candidate's score.
  </Accordion>
</AccordionGroup>

**An off-script question can strengthen a competency. It can never create one.** Weights are set by you before the candidate is known, and no interviewer can change them mid-round.

## Coverage is reported, not deducted — where the gap was not the candidate's

The guiding rule: **a candidate is scored only on what they controlled.**

| Round type                  | If questions go unasked  | Why                                                                    |
| --------------------------- | ------------------------ | ---------------------------------------------------------------------- |
| **Panel and HR**            | Reported, never deducted | The interviewer chose what to ask. The candidate had no say.           |
| **AI voice, avatar, phone** | Deducted                 | The agent asks the full list, so a gap means the candidate disengaged. |

You will always see how much of your question list was covered. On a human-led round, that figure informs you — it does not penalise the candidate for their interviewer's choices.

## What a thin round can conclude

A round's completeness is measured by **how much of your competency matrix was assessed** — not how many questions were asked. Those are very different things:

* 20% of your list asked, but conversation covered every competency → **complete assessment**
* 60% of your list asked, all of it inside one competency → **incomplete assessment**

| Level              | What was assessed                        | What the report can recommend                           |
| ------------------ | ---------------------------------------- | ------------------------------------------------------- |
| **Conclusive**     | 80%+ of matrix weight, 5+ scored answers | Full recommendation                                     |
| **Provisional**    | 50%+ of matrix weight, 3+ scored answers | Full recommendation, with unassessed competencies named |
| **Indicative**     | Below that                               | **Hold only** — cannot advance or reject                |
| **Not assessable** | No scored answers                        | No recommendation; the round is returned for review     |

The limit on an Indicative round runs **both ways**. One answer at 10/10 is not grounds to advance someone, and one answer at 0/10 is not grounds to reject them. A thin round never becomes a one-way argument against a candidate.

The score itself is never adjusted for thinness — only what may be concluded from it.

## Score bands

| Score    | Band   | Recommendation |
| -------- | ------ | -------------- |
| 80–100   | High   | Proceed        |
| 60–79    | Medium | Proceed        |
| 40–59    | Low    | Hold           |
| Below 40 | Reject | Reject         |

Every round type uses this same table. A 72 on a panel round means what a 72 means on an AI voice round.

## The guarantees

<CardGroup cols={2}>
  <Card title="Reproducible" icon="rotate">
    The same answers always produce the same score. No randomness, and scoring
    models run at zero temperature.
  </Card>

  <Card title="Recomputable" icon="calculator">
    Every report stores the competency scores and the weights applied, so the
    headline can be re-derived by hand at any time.
  </Card>

  <Card title="Version-pinned" icon="lock">
    A score is pinned to the weights and method in force when it was produced.
    Changing your weights never rewrites a score someone already acted on.
  </Card>

  <Card title="Never model-declared" icon="shield-check">
    The AI scores individual answers. Exterview's servers do the weighting and
    produce the headline — the model never emits the number a decision is made
    on.
  </Card>
</CardGroup>

## What this does not claim

Being straight about the limits is part of being trustworthy about the method.

* **This makes aggregation fair. It does not make the underlying judgement perfect.** The 0–10 score on each answer is still an AI assessment of that answer.
* **Scoring depends on the transcript being correct.** If a transcript misattributes who said what, the scores will be confidently wrong.
* **A score is a recommendation, never a decision.** Every outcome is subject to human review, and no adverse decision is issued automatically from an incomplete round.

## FAQs

<AccordionGroup>
  <Accordion title="Why did a candidate score lower than the individual answers suggest?">
    Most often because a competency you weight heavily was assessed weakly, or
    barely at all. Check the per-competency table and the unassessed list — the
    headline follows your weighting, not the raw number of good answers.
  </Accordion>

  <Accordion title="Can I recompute a score myself?">
    Yes, and you are encouraged to. Take each competency's average, multiply by
    its weight, add them up, divide by the total weight of the competencies that
    were actually assessed, multiply by 100.
  </Accordion>

  <Accordion title="Does changing my weights change past scores?">
    No. Existing reports keep the weights they were scored with. New weights
    apply to new rounds.
  </Accordion>

  <Accordion title="Does a candidate lose points for questions the interviewer skipped?">
    Not on a panel or HR round. Coverage is reported so you can see what was
    missed, but it does not reduce the score.
  </Accordion>

  <Accordion title="What can we tell a candidate who asks why they were scored the way they were?">
    The competencies assessed, their weights, the answers behind each competency
    score, and how much of the matrix was covered — all of it is on the report,
    and all of it is the actual basis for the number.
  </Accordion>
</AccordionGroup>

## Related

* [Report Glossary](/monitor/report-glossary) — every report type and score dimension defined
* [How to Set Scoring Weights](/guides/setup-scoring) — configuring your competency mix
* [Scoring Quality](/best-practices/scoring-quality) — getting better signal out of rounds
* [Calibration](/monitor/calibration) — keeping scores aligned to your outcomes