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

# Accuracy Guarantee

> Our validation methodology and quality commitment

# Our Commitment to Accuracy

Legal documents demand precision. A single incorrect change can have
significant consequences. That's why we've built the most rigorous
validation system in the industry.

## The Accuracy Problem

Traditional document editing has risks:

<Warning>**Hallucination**: AI adding content that wasn't requested</Warning>

<Warning>**Scope Creep**: Changes extending beyond specified sections</Warning>

<Warning>
  **Format Corruption**: Broken formatting, numbering, or layout
</Warning>

<Warning>**Silent Errors**: Issues that aren't visible until signing</Warning>

## Our Solution

We solve these problems through multi-layer validation:

```mermaid theme={null}
graph TD
    A[Document Edited] --> B[Pass 1: Content Check]
    B --> C[Pass 2: Risk Check]
    C --> D[Pass 3: Scope Check]
    D --> E[Pass 4: Format Check]
    E --> F[Pass 5: Visual Check]
    F --> G{All Passed?}
    G -->|No| H[Refine & Retry]
    H --> A
    G -->|Yes| I[Deliver]
```

***

## Change Categorization

Every modification is categorized:

### ✅ Requested Changes

Changes that directly match your instructions. These are expected and approved.

> You said: "Change Delaware to Texas"
> We changed: "Delaware" → "Texas" in governing law clause

### ✅ Derived Changes

Changes that logically follow from your instructions but weren't explicitly
stated. These are noted and approved.

> You said: "Make the NDA mutual"
> We also updated: pronouns, headings, and cross-references for consistency

### ⚠️ Unexpected Changes

Changes that can't be traced to your instructions. These are **flagged for
review** before delivery.

> Not requested, but found: Additional text in Section 5
> Status: Flagged, pending your review

### 🚨 Boilerplate Detection

Generic legal text that appeared from nowhere is **blocked**. This is the
strongest indicator of AI hallucination.

> Detected: Standard confidentiality boilerplate not in original
> Status: Blocked, not included in output

***

## Validation Passes Explained

### Pass 1: Content Accuracy

* Do changes match instructions?
* Is anything missing that was requested?
* Is anything present that wasn't requested?
* **Blocks**: Unexpected or boilerplate content

### Pass 2: Risk Analysis

* Are there high-risk legal changes?
* Were liability or indemnification terms modified?
* Are any changes potentially harmful?
* **Flags**: High-risk changes for review

### Pass 3: Scope Compliance

* Did changes stay within specified sections?
* Were only targeted areas modified?
* Is the rest of the document untouched?
* **Flags**: Out-of-scope modifications

### Pass 4: Format Integrity

* Is document structure preserved?
* Is numbering sequential?
* Are there any format artifacts?
* **Blocks**: Corrupt or injected formatting

### Pass 5: Visual Verification

* Does the document look correct when rendered?
* Are there any visual anomalies?
* Would a human reviewer notice issues?
* **Flags**: Visual problems for attention

***

## Automatic Retry

When validation fails, we don't just report the error. We:

<Steps>
  <Step title="Identify the Issue">
    What exactly went wrong? Missing content? Extra content? Wrong section?
  </Step>

  <Step title="Refine Approach">
    Adjust our process to address the specific issue.
  </Step>

  <Step title="Try Again">Re-apply changes with refined approach.</Step>
  <Step title="Re-validate">Run all validation passes again.</Step>
</Steps>

This continues up to three times. You receive the best possible result.

***

## What You Can Expect

<CardGroup cols={2}>
  <Card title="99%+ Accuracy" icon="bullseye">
    Requested changes are applied correctly over 99% of the time.
  </Card>

  <Card title="Zero Hallucinations" icon="shield-check">
    Boilerplate injection is blocked. You get what you asked for.
  </Card>

  <Card title="Full Transparency" icon="eye">
    Every change is categorized and visible. No surprises.
  </Card>

  <Card title="Audit Trail" icon="scroll">
    Complete history of what was done and why.
  </Card>
</CardGroup>

***

## When Human Review Is Needed

Some situations warrant human attention:

<Info>
  **High-Risk Changes**: Modifications to indemnification, liability, or
  termination clauses are flagged for senior review.
</Info>

<Info>
  **Unexpected Changes**: Any change not traceable to instructions is surfaced
  for your decision.
</Info>

<Info>
  **Low Confidence**: If we're not sure about something, we ask rather than
  guess.
</Info>

***

## Comparison with Alternatives

| Aspect                   | Basic AI Editing | mixus Legal AI         |
| ------------------------ | ---------------- | ---------------------- |
| Hallucination prevention | ❌ No             | ✅ 4-layer protection   |
| Change categorization    | ❌ No             | ✅ Every change tracked |
| Visual verification      | ❌ No             | ✅ PDF rendering check  |
| Automatic retry          | ❌ No             | ✅ Up to 3 attempts     |
| Risk detection           | ❌ No             | ✅ Automatic flagging   |
| Audit trail              | ❌ No             | ✅ Complete history     |

***

## Our Guarantee

If you find an error in our output that wasn't flagged, we want to know.
Our system improves with feedback, and we're committed to the highest
standards of accuracy.

<Card title="Enterprise Features" icon="building" href="/legal-ai/enterprise-features">
  Learn about audit trails and compliance features for enterprise use.
</Card>