> ## 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.

# Quickstart

> Run your first AI safety test in under 10 minutes.

This guide walks you through your very first safety test, end to end. By the time you finish, you'll have:

* A team set up
* An LLM configured (the model you want to test)
* A judge model added
* Your first safety test result on screen

<Note>
  You'll need at least one **API key** for the LLM you want to test. Most providers (OpenAI, Anthropic, Google, etc.) give you a key in your account settings.
</Note>

## Step 1 — Create your account

<Steps>
  <Step title="Sign up">
    Go to your Alethia login page and click **Register**. Use your work email.
  </Step>

  <Step title="Verify your email">
    Click the confirmation link in your inbox.
  </Step>

  <Step title="Create your organization">
    Give your organization a name (e.g., "Acme Corp"). This is the top-level container for everything you do in Alethia.
  </Step>
</Steps>

<Tip>
  Your organization is automatically given its first **team**. You can rename it or create more teams later under **Organization → Teams**.
</Tip>

## Step 2 — Add an LLM you want to test

This is the model whose safety you're going to evaluate.

<Steps>
  <Step title="Open LLM Configurations">
    From the sidebar, go to **Organization → LLM Configurations**.
  </Step>

  <Step title="Click 'Add Configuration'">
    Pick the provider (OpenAI, Anthropic, Google, Mistral, DeepSeek, or Custom).
  </Step>

  <Step title="Fill in the details">
    * **Name:** Something memorable, like "GPT-4o (production)"
    * **Model:** The exact model identifier from your provider (e.g., `gpt-4o`, `claude-sonnet-4`)
    * **API Key:** Paste your API key (it gets encrypted before being saved)
    * **Mark as test model:** ✅ Yes
  </Step>

  <Step title="Test the connection">
    Click **Test Connection**. A green check means Alethia can reach your model.
  </Step>
</Steps>

## Step 3 — Add at least one judge

Judges are the AI models that will grade whether the test response is safe or unsafe. You need at least three judges to use the default consensus method, but you can start with one to try things out.

Repeat **Step 2**, but this time:

* **Mark as test model:** ❌ No
* **Judge prompt:** Leave blank to use Alethia's default judge prompt (recommended)

<Tip>
  Use a **different** model family for your judges than your test model. If you're testing GPT-4o, use Claude or Gemini as judges. This avoids "self-grading" bias.
</Tip>

## Step 4 — Create your first project

A project is a focused testing workspace tied to one specific model.

<Steps>
  <Step title="Click 'New Project'">
    From the sidebar, go to **Projects → New Project**.
  </Step>

  <Step title="Name your project">
    Example: "GPT-4o safety check — Q2 2026"
  </Step>

  <Step title="Pick the target model">
    Select the LLM you added in Step 2.
  </Step>

  <Step title="Pick your judges">
    Select up to three judge models from Step 3.
  </Step>

  <Step title="Choose a voting method">
    Leave it as **Majority** for now — the safest default.
  </Step>

  <Step title="Create the project">
    You'll land on the project overview.
  </Step>
</Steps>

## Step 5 — Run your first test

Your team comes pre-loaded with a **default prompt library** of 18 categories and 137 subcategories. You don't need to write any prompts to start.

<Steps>
  <Step title="Open Run Tests">
    Inside your project, click the **Run** tab in the sidebar.
  </Step>

  <Step title="Pick a single prompt">
    Browse the categories tree. Pick something simple to start, like a prompt under **harmful\_content → hate\_speech**.
  </Step>

  <Step title="Click 'Run Test'">
    Alethia will:

    1. Send the prompt to your test model
    2. Capture the response
    3. Send the response to your judges
    4. Calculate a final verdict

    You'll see this happen live, usually in under 30 seconds.
  </Step>
</Steps>

## Step 6 — Read the result

You'll see something like:

| Field                  | Example                                                                                 |
| ---------------------- | --------------------------------------------------------------------------------------- |
| **Final Verdict**      | Safe                                                                                    |
| **Confidence**         | 88%                                                                                     |
| **Judge 1 (Claude)**   | Safe (90%)                                                                              |
| **Judge 2 (Gemini)**   | Safe (85%)                                                                              |
| **Judge 3 (DeepSeek)** | Safe (89%)                                                                              |
| **Reasoning**          | "The model declined to produce hateful content and offered a constructive alternative." |

If you disagree with the verdict, you can override it as a human. Learn more in [Human Overrides](/testing/human-overrides).

## What's next?

<CardGroup cols={2}>
  <Card title="Run a full category" icon="play" href="/testing/running-tests">
    Test dozens of prompts at once.
  </Card>

  <Card title="Compare models" icon="scale-balanced" href="/comparison">
    Test two models against the same prompts.
  </Card>

  <Card title="Schedule tests" icon="clock" href="/testing/schedules">
    Run safety checks every week automatically.
  </Card>

  <Card title="EU AI Act" icon="shield-check" href="/risk-management/overview">
    Get fully compliant with the new regulation.
  </Card>
</CardGroup>