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

# Quality

> How agent quality is scored, what the reports show, and how corrections feed back into the agent.

Every conversation is scored. Quality reporting is where you see whether the agent is actually good,
and where it is not.

## The tabs

| Tab                     | Shows                                                             |
| ----------------------- | ----------------------------------------------------------------- |
| **Overview**            | QA scores, hallucination flag rates, trends                       |
| **Escalations**         | How often the agent hands off and why                             |
| **[SLA](/inbox/sla)**   | Hit rate, median, 90th percentile, breaches                       |
| **[CSAT](/inbox/csat)** | What customers actually said                                      |
| **Assistant**           | How your [team assistants](/assistants/governance) are performing |

## How scoring works

Every reply is graded on the live path for unsupported claims, with no model call, so it costs no
latency. See [Guardrails](/agent/guardrails).

Beyond that, conversations are evaluated against a rubric covering whether the customer's question
was answered, whether the agent followed your rules, tone and persona fidelity, and whether the
conversation reached a sensible outcome.

Scores are aggregated so you see distributions rather than a single blended number. A workspace with
a 4.2 average made up of mostly 5s and a tail of 1s has a different problem from one that is
uniformly mediocre.

## Signals that feed quality

<CardGroup cols={2}>
  <Card title="Automatic grading" icon="robot">
    Every turn, on the live path plus an asynchronous deeper pass.
  </Card>

  <Card title="Operator feedback" icon="thumbs-up">
    Thumbs up and down from your team, with a reason. See [Explainability](/agent/explainability).
  </Card>

  <Card title="Customer ratings" icon="star">
    CSAT, which is the only signal that comes from outside.
  </Card>

  <Card title="Outcomes" icon="flag-checkered">
    Whether conversations resolved, escalated, or were abandoned.
  </Card>
</CardGroup>

## Calibration

The platform also tracks whether the agent's confidence matches its accuracy. An agent that hedges
when it is right and asserts when it is wrong is poorly calibrated even with an acceptable
hallucination rate, and it erodes trust faster than occasional errors do.

## Closing the loop

Quality data is not just a report.

* **Failures feed [knowledge gaps](/agent/knowledge).** A cluster of bad answers on one topic becomes
  a ranked, actionable finding.
* **Corrections teach [categories](/inbox/categories).** Fixing a wrong category is training data.
* **Prompt changes can be tested rather than guessed.** Run two variants against real traffic on a
  split and compare their scores before committing.
* **Releases can be compared and rolled back.** See [The agent](/agent/overview).

## Before you go live and after

Prompt and configuration changes are evaluated against recorded conversations before they reach
customers, so a change that would have made past conversations worse is caught first. This is how a
change gets promoted rather than merely deployed.

## Reading the numbers honestly

* **Look at the distribution, not the average.**
* **Compare like with like.** Channel mix and volume changes move quality numbers on their own.
* **Where CSAT and QA disagree, read the conversation.** The platform's opinion of its own work and
  the customer's opinion diverging is the most informative thing in the report.
* **A rising handoff rate is not automatically bad.** An agent that escalates more because your rules
  got stricter is working correctly.