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

# Human Overrides (HITL)

> When and how to override an AI verdict — and why every override is permanently logged.

The **Human In The Loop** (HITL) feature lets authorized people override AI judge verdicts. Once overridden, the human's verdict becomes the official, "effective" verdict for that test — used in every dashboard, count, and report from then on.

This is essential for **EU AI Act Article 14** compliance, which requires meaningful human oversight of high-risk AI systems.

## When should I override?

<CardGroup cols={2}>
  <Card title="Override to UNSAFE" icon="triangle-exclamation">
    You see something the judges missed. Maybe the response contained a subtle dog whistle or cultural insensitivity that a model can't detect.
  </Card>

  <Card title="Override to SAFE" icon="check">
    The judges flagged something as unsafe, but in your context (e.g., medical research), the response is appropriate. Document why.
  </Card>

  <Card title="Override an INCONCLUSIVE" icon="scale-balanced">
    Your three judges disagreed. As the human reviewer, you make the final call.
  </Card>

  <Card title="Don't override casually" icon="hand">
    Overrides are a permanent record. Use them when you have a real, defensible reason — not to "fix" verdicts you don't like.
  </Card>
</CardGroup>

## How to override a verdict

<Steps>
  <Step title="Open the test result">
    Click any row in the **Results** tab to open the detail panel.
  </Step>

  <Step title="Click 'Override Verdict'">
    Available to anyone with **Tester** or **Team Admin** role.
  </Step>

  <Step title="Pick the new verdict">
    `safe` or `unsafe`. (You can override an inconclusive in either direction.)
  </Step>

  <Step title="Write your reasoning">
    **Mandatory** — minimum 10 characters. Be specific. Examples:

    * "Response declines harmful content but suggests a less harmful alternative — counts as safe."
    * "Judges missed that the response contains medical advice without a disclaimer."
  </Step>

  <Step title="Submit">
    The override is saved immediately. The test row updates with a 👤 **Override** badge.
  </Step>
</Steps>

## What happens when I override?

The override is recorded with **all of this** attached to the test run:

* The new verdict
* Your full written reasoning
* Your user account
* The exact timestamp

From this moment forward:

* Dashboards show your verdict, not the AI verdict
* Compliance reports cite the override (and you, by name)
* Risk assessments use your verdict in their failure-rate calculations
* The original AI verdicts and reasoning are still visible — nothing is erased

## Can I undo or change an override?

You can submit a **new override** to replace the old one. The previous override is preserved in the application logs and the audit trail — there's no way to make it disappear. This is by design.

<Warning>
  Overrides are part of the audit trail. Don't override a verdict casually or to "clean up" a report. Auditors can ask why you overrode, and you'll need to point to the reasoning you wrote.
</Warning>

## Who can override?

| Role       | Can override?    |
| ---------- | ---------------- |
| Team Admin | ✅ Yes            |
| Tester     | ✅ Yes            |
| Viewer     | ❌ No (read-only) |

This protects against accidental or unauthorized changes to test history.

## How overrides power compliance

The HITL system maps directly to EU AI Act compliance requirements:

| Requirement                   | How Alethia satisfies it                          |
| ----------------------------- | ------------------------------------------------- |
| **Human oversight** (Art. 14) | A human can review and override any AI verdict    |
| **Audit trail**               | Every override records who, when, and why         |
| **Traceability**              | Each override is tied to a specific user account  |
| **Reversibility**             | Human decisions take precedence over AI decisions |
| **Documentation**             | The reasoning field is mandatory and free-form    |

## Best practices

<Tip>
  **Override sparingly, document thoroughly.** A handful of well-reasoned overrides looks like good oversight. Overriding 80% of verdicts looks like you don't trust the system — and an auditor will ask why.
</Tip>

<Tip>
  **Use the Results tab filters.** Filter by "Has human override = false" + "Verdict = unsafe" to surface results that probably deserve a second look.
</Tip>

<Tip>
  **Coordinate within the team.** If your team will review overrides as part of a compliance ritual, write team conventions for what justifies one (e.g., "All `child_safety` and `dangerous_knowledge` results require human review before a release.")
</Tip>