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

# Rubrics

> How intervyo.ai scores interviews — dimensions, weights, thresholds, calibration.

A **rubric** is the structured scoring criteria attached to an evaluation
template. It's how intervyo.ai turns a free-form voice interview into a
defensible scorecard.

Rubrics are versioned, calibrated by default, and applied identically to
every participant — same dimensions, same weights, same bar.

## Anatomy of a rubric

<CodeGroup>
  ```json Structure theme={null}
  {
    "dimensions": [
      {
        "id": "problem_decomposition",
        "label": "Problem decomposition",
        "weight": 0.30,
        "passThreshold": 6.5,
        "description": "Can the candidate break ambiguity into concrete sub-problems and pick a place to start?"
      },
      {
        "id": "code_quality",
        "label": "Code quality & correctness",
        "weight": 0.30,
        "passThreshold": 7.0
      },
      {
        "id": "tradeoff_thinking",
        "label": "Trade-off thinking",
        "weight": 0.25,
        "passThreshold": 6.0
      },
      {
        "id": "communication",
        "label": "Communication under pressure",
        "weight": 0.15,
        "passThreshold": 6.0
      }
    ],
    "overallPassThreshold": 7.0
  }
  ```
</CodeGroup>

<ParamField path="dimensions[].weight" type="number">
  Fractional weight in the overall score. Weights across all dimensions
  should sum to `1.0`. The platform normalizes if they don't, but explicit
  is better.
</ParamField>

<ParamField path="dimensions[].passThreshold" type="number">
  Minimum score on this dimension for a pass. A candidate that scores
  below the threshold on **any single dimension** fails the rubric — even
  if their overall is above `overallPassThreshold`.
</ParamField>

<ParamField path="overallPassThreshold" type="number">
  Minimum weighted-average score for a pass. Defaults to `7.0` if unset.
</ParamField>

## How scoring runs

For each dimension, the AI:

<Steps>
  <Step title="Reads the transcript">
    Identifies the segments relevant to this dimension — code-quality
    evidence in the coding round, discovery-question evidence in a sales
    round.
  </Step>

  <Step title="Grades against the dimension's description">
    Produces a 0–10 score with a short reasoning paragraph citing the
    specific transcript lines that drove the score.
  </Step>

  <Step title="Joins into the breakdown">
    All per-dimension scores are bundled into `evaluation_breakdown` with
    transcript citations and reasoning per dimension.
  </Step>

  <Step title="Computes overall + pass">
    Weighted average → overall score. Compare against thresholds → emit
    `passed: true|false`.
  </Step>
</Steps>

## Why dimensions matter more than overall

A single overall score throws away the signal you need to actually decide.
Two candidates can both score 7.5 overall and be wildly different hires:

| Candidate | Communication | Technical | Trade-offs | Authenticity |
| --------- | ------------- | --------- | ---------- | ------------ |
| **Alex**  | 9.0           | 6.5       | 7.5        | 9.0          |
| **Brin**  | 6.0           | 9.0       | 7.5        | 9.0          |

Both score 7.5. Alex is a strong PM-leaning IC; Brin is a strong heads-down
engineer. **The dimension breakdown is the hiring signal**, not the
average.

<Tip>
  Show your hiring panel the per-dimension breakdown, not the overall.
  Decisions converge faster when everyone's looking at the same evidence
  per dimension, with the transcript citations one click away.
</Tip>

## Templates ship with calibrated rubrics

You don't have to design rubrics from scratch. intervyo.ai ships
calibrated starters for the most-hired roles:

<CardGroup cols={2}>
  <Card title="Software engineers" icon="code">
    Problem decomposition · Code quality · Trade-off thinking · Communication
  </Card>

  <Card title="Sales reps" icon="trending-up">
    Discovery quality · Pitch adaptation · Objection handling · Closing
  </Card>

  <Card title="Customer support" icon="life-buoy">
    Diagnosis · Written communication · Technical fluency · Escalation judgement
  </Card>

  <Card title="Product managers" icon="target">
    Prioritization · Customer empathy · Cross-functional language · Metric literacy
  </Card>
</CardGroup>

See the [full list of role templates](https://intervyo.ai/solutions/by-role).
You can use them as-is, or fork them — every starter is a calibrated
baseline, not a contract.

## Calibration analytics

Rubrics are useful only if humans use them consistently. intervyo.ai
surfaces:

* **Recruiter-to-recruiter variance** per dimension
* **Hiring-manager parity** across panels reviewing the same candidates
* **Drift alerts** when a recruiter's scoring trends away from the team
  baseline by more than a threshold
* **Calibration sessions** — built-in quarterly workflow where multiple
  scorers grade the same candidate and the platform surfaces divergence

Read more under [Recruiter Calibration](https://intervyo.ai/features/recruiter-calibration)
on the marketing site.

## Versioning

Rubrics on a template are **versioned**. Edits create a new version;
existing sessions keep the rubric they were scored against. That way:

* Hiring panels reviewing 60-day-old scorecards see the rubric that
  produced those scores
* You can tighten a rubric mid-quarter without breaking historical
  comparability
* Compliance teams can audit which version of which rubric scored which
  candidate

<Warning>
  Don't tweak a rubric and run a back-comparison expecting like-for-like
  numbers. A 7.5 under v3 of the rubric is not directly comparable to a
  7.5 under v2 — the dimensions or weights changed.
</Warning>