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

# How DualMind Works

> Understand the core mechanics of DualMind Arena — from blind model comparison to ELO scoring.

## The Blind Comparison Concept

The most effective way to evaluate AI models is to remove brand bias. If you know a response comes from "GPT-4" or "Claude 3.5 Sonnet", you are subconsciously primed to rate it higher.

DualMind Arena solves this with **Blind Battles**:

1. **Submit a Prompt**: You send a single prompt to the arena.
2. **Anonymous Generation**: Two different models generate responses simultaneously. Labels are hidden (e.g., "Model A" vs "Model B").
3. **Vote**: You select the better response based purely on quality, accuracy, and helpfulness.
4. **Reveal**: Only *after* you vote are the model identities revealed.

This methodology creates a dataset of pure quality preference, unpolluted by marketing or brand reputation.

## The ELO Rating System

We use the **ELO rating system** — the same system used in Chess and competitive video games — to rank AI models.

* **Starting Score**: All models start with a baseline rating (e.g., 1000).
* **Winning**: Beating a high-rated model awards more points than beating a low-rated one.
* **Losing**: Losing to a low-rated model costs more points than losing to a highly-rated champion.
* **Ties**: Points are distributed based on the rating difference (a lower-rated model drawing with a higher-rated one gains points).

This system is self-correcting. Over thousands of battles, it produces a highly accurate hierarchy of model capability that reflects real-world usage patterns rather than abstract benchmarks.

## Data Processing Pipeline

When you submit a prompt to DualMind Arena, our platform orchestrates a complex evaluation pipeline in milliseconds:

<Steps>
  <Step title="Safety & Moderation">
    Incoming prompts are scanned for safety policy compliance to ensuring the arena remains a constructive environment.
  </Step>

  <Step title="Model Orchestration">
    The system selects model pairs based on your chosen mode (Random, Topper, or Manual) and routes the request to the appropriate inference providers.
  </Step>

  <Step title="Parallel Inference">
    Both models process the prompt simultaneously. We normalize response times to ensure speed differences don't bias your voting decision (unless speed is your specific criteria).
  </Step>

  <Step title="Response Normalization">
    Markdown formatting, code blocks, and LaTeX math are standardized to ensure visual consistency between different models' outputs.
  </Step>
</Steps>

## Platform Architecture

DualMind is built for high availability and low latency.

* **Global Edge Network**: Our frontend is served from edge locations worldwide to minimize initial load time.
* **Provider Resilience**: We integrate with multiple AI inference providers. If one provider experiences downtime, traffic is automatically rerouted to ensure the arena remains active.
* **Live Leaderboards**: Voting data is processed in real-time, meaning the leaderboard you see always reflects the very latest community consensus.