> ## 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.

# Human Review and Reading AI Reports

> How to keep hiring decisions human-owned and defensible in Exterview: mandatory review, how to read a report well, and when a step should gate versus inform.

## Purpose

In regulated hiring, a decision has to be defensible: a person owned it, the reasoning is recorded, and the evidence is reviewable. Exterview is built so the AI sharpens that decision without ever making it. This page covers the standard, how to read a report well, and how to decide whether a step should block progression or just inform it.

## The standard

<CardGroup cols={2}>
  <Card title="A person decides">
    No adverse decision is made autonomously. The AI proposes and evidences; a person decides.
  </Card>

  <Card title="Reasoning is recorded">
    The why behind a decision is captured, not just the outcome.
  </Card>

  <Card title="Evidence is reviewable">
    Every score traces to session evidence and the configuration version that
    produced it.
  </Card>

  <Card title="Content, not biometrics">
    Scoring is grounded in what a candidate says, never facial analysis or tone.
  </Card>
</CardGroup>

## How to read a report well

An Exterview report is decision support, not a verdict. Reviewing it well means reading the evidence behind the score, not just the number.

<Steps>
  <Step title="Start with the evidence, not the score">
    Read what the candidate actually said and how it maps to the rubric before
    you look at the number. The score summarizes the evidence; it does not
    replace it.
  </Step>

  <Step title="Check the citations">
    Every claim traces to a moment in the session. Spot-check the citations that
    most affect the decision.
  </Step>

  <Step title="Look for a critic flag">
    If the critic layer flagged the output, read the flag first once that
    capability is available. See [AI Observability](/monitor/ai-observability).
  </Step>

  <Step title="Make the call">
    You own the decision. Use the report to make it faster and more defensible,
    not to outsource it.
  </Step>
</Steps>

### Good habits

<CardGroup cols={2}>
  <Card title="Read across candidates">
    Compare on the same rubric dimensions rather than on overall impression.
  </Card>

  <Card title="Weight gated dimensions">
    Give the most attention to the dimensions your configuration treats as gates.
  </Card>

  <Card title="Record the why">
    Note the reason for the decision. It feeds auditability and the post-hire
    loop.
  </Card>

  <Card title="Assign clear reviewers">
    Every gate has a named human owner. Configure this in [Governance](/configure/governance).
  </Card>
</CardGroup>

## Gate vs. signal: what should block progression

Every evaluation step in Exterview is configured as a gate or a signal. Getting this split right is what keeps your pipeline both fair and efficient. Over-gate and you reject good candidates on one weak dimension; under-gate and nothing is actually filtered. See [Operations](/configure/operations) for where this is set.

| Type       | What it does                                                         | Use it when                                                                     |
| ---------- | -------------------------------------------------------------------- | ------------------------------------------------------------------------------- |
| **Gate**   | A pass or fail checkpoint that controls whether a candidate advances | The dimension is a genuine must-have where a low result should stop progression |
| **Signal** | An input that informs the decision without blocking it               | The dimension is important context but not independently disqualifying          |

<Steps>
  <Step title="Ask if it is truly disqualifying">
    If a weak result on this dimension alone should end the process, it is a
    gate. If not, it is a signal.
  </Step>

  <Step title="Default to signal">
    When unsure, make it a signal. Signals inform a human decision; gates remove
    it. Reserve gates for real must-haves.
  </Step>

  <Step title="Limit the number of gates">
    A few well-chosen gates filter effectively. Many gates compound into a
    funnel that rejects strong candidates on a single off dimension.
  </Step>

  <Step title="Keep gates human-reviewed">
    A gate controls progression, but the decision at the gate is still owned by
    a person, not an autonomous reject.
  </Step>
</Steps>

### Common mistakes

<CardGroup cols={2}>
  <Card title="Over-gating">
    Turning every dimension into a gate rejects good candidates on one weak area.
  </Card>

  <Card title="Gating soft skills rigidly">
    Nuanced dimensions usually belong as signals a human weighs, not hard gates.
  </Card>

  <Card title="No gates at all">
    If nothing gates, the pipeline filters nothing and reviewers drown.
  </Card>

  <Card title="Silent auto-reject">
    A gate should route to human review, never auto-reject without a person.
  </Card>
</CardGroup>

<Warning>
  Never configure a fully automated reject, and do not rubber-stamp the
  top-ranked candidate without reading the evidence. Exterview does not make
  autonomous adverse decisions, and your process should not reintroduce one
  around it.
</Warning>

## Related

<CardGroup cols={2}>
  <Card title="Responsible AI" href="/trust/responsible-ai">
    The platform-wide human-review standard.
  </Card>

  <Card title="Auditability" href="/trust/auditability">
    The traceable record behind every decision.
  </Card>

  <Card title="Evaluation Design" href="/best-practices/evaluation-design">
    Mapping competencies to the right steps.
  </Card>

  <Card title="Scoring Quality" href="/best-practices/scoring-quality">
    Keeping scores consistent across reviewers.
  </Card>
</CardGroup>