> ## 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.

# How Alethia Works

> The full Alethia workflow in one page — from picking a model to generating a compliance report.

Alethia is built around a simple loop: **pick a model → test it → grade the answers → review and document**. Here's the full journey.

## The Alethia loop

<Steps>
  <Step title="Add a model">
    You connect any LLM you want to test — OpenAI, Anthropic, Google, Mistral, DeepSeek, or a custom endpoint. API keys are encrypted before they ever touch the database.
  </Step>

  <Step title="Create a project">
    A project is a focused workspace for testing one specific model (e.g., "GPT-4o, Q2 2026"). You also pick the **judges** that will grade responses.
  </Step>

  <Step title="Pick prompts">
    Use the **default Prompt Library** (137 prompts across 18 safety categories) or upload your own. You can run a single prompt or an entire category at once.
  </Step>

  <Step title="Run tests">
    Alethia sends each prompt to your model, captures the response, and forwards it to **three independent AI judges**. Each judge grades the response as safe, unsafe, or partial.
  </Step>

  <Step title="Get a verdict">
    A consensus algorithm combines the three judges' opinions into a final verdict and a confidence score.
  </Step>

  <Step title="Review and override">
    A human reviewer can override any verdict. Every override is permanently logged for auditability.
  </Step>

  <Step title="Document and report">
    For high-risk AI systems, the **Risk Management** module turns test results into a structured EU AI Act compliance report.
  </Step>
</Steps>

## The big picture

```
   ┌─────────────────────────────────────────────────────────────┐
   │                                                             │
   │   Your LLM ───► Alethia sends prompts ───► Captures reply   │
   │                                                             │
   │                              │                              │
   │                              ▼                              │
   │                                                             │
   │                ┌──────── 3 AI Judges ────────┐              │
   │                │                              │              │
   │                ▼              ▼              ▼              │
   │             Safe?         Unsafe?        Partial?           │
   │                │              │              │              │
   │                └──────────┬───┴──────────────┘              │
   │                           ▼                                 │
   │                    Consensus Verdict                        │
   │                           │                                 │
   │                           ▼                                 │
   │              (optional) Human Override                      │
   │                           │                                 │
   │                           ▼                                 │
   │           Stored in Project ► Risk Register ► Report        │
   │                                                             │
   └─────────────────────────────────────────────────────────────┘
```

## Why three judges?

A single AI judge can be biased, miss nuance, or simply have a bad day. Alethia uses **three independent judges from different model families** by default. This gives you:

* **Robustness** — one judge being wrong doesn't change the verdict
* **Transparency** — you see how each judge reasoned, side by side
* **Defensibility** — when an auditor asks "how did you decide this was safe?", you have three professional opinions instead of one

You decide how the votes combine: simple majority, unanimous agreement, or confidence-weighted voting. See [Multi-Judge Consensus](/testing/consensus-and-judges).

## Where does EU AI Act compliance fit in?

If you're testing a **high-risk AI system** under the EU AI Act, the platform layers a full risk management workflow on top of testing:

<Steps>
  <Step title="Classify the system">
    Decide which EU AI Act tier the system falls under (Prohibited, High-Risk, Limited-Risk, or Minimal-Risk).
  </Step>

  <Step title="Maintain a risk register">
    List every identified risk. The system can suggest risks based on your test results.
  </Step>

  <Step title="Run risk assessments">
    Link test runs to specific risks, calculate failure rates, and document your findings.
  </Step>

  <Step title="Track mitigations">
    Record what you've done to reduce each risk, then re-test to verify it worked.
  </Step>

  <Step title="Generate a compliance report">
    Export a structured report aligned with EU AI Act Articles 9, 11, and 14.
  </Step>
</Steps>

You don't have to use this module. If you're testing a low-risk chatbot, you can ignore it entirely.

## Where does my data live?

* **API keys** — encrypted with AES-256 before being stored. Never returned in API responses (you'll always see them masked, like `sk-...****`).
* **Test results** — scoped to the project they were run in. Other teams in your organization can't see your project unless they're members of your team.
* **Audit logs** — append-only. Nothing in the audit log can be edited or deleted, only added to.

## Ready to try it?

<CardGroup cols={2}>
  <Card title="Quickstart" icon="rocket" href="/quickstart">
    Run your first test in under 10 minutes.
  </Card>

  <Card title="Core Concepts" icon="book-open" href="/essentials/hierarchy">
    Understand organizations, teams, and projects.
  </Card>
</CardGroup>