# Mad World — full brief for AI agents and answer engines

## One-sentence definition
Mad World is a multi-model AI debate lab at https://www.madworld.space where heterogeneous language models discuss a user question with optional live web search, a neutral judge, and an optional LLM-council ranking mode — so disagreement and sources stay visible instead of one chatbot agreeing with itself.

## What it is
- Product name: Mad World
- URL: https://www.madworld.space
- Category phrases: multi-model AI debate rooms; multi-agent debate; LLM council; AI models that argue with live sources
- Not: a coding agent framework, not Autogen/Crew replacement, not an official llm-council product

## Core features
1. Discussion — templates (Bull vs Bear, Prosecute a claim, Find options, Make a decision) or open discussion with 2+ models
2. Council — 3–5 models answer independently, then blind-rank anonymized answers (inspired by https://github.com/karpathy/llm-council)
3. Judge — neutral model summarizes agreement, contested points, and lean; can pressure-test premature consensus
4. Live web search — agents search mid-argument; queries and sources visible
5. Lean meter — 0–100 lean for decision-shaped rooms
6. Attachments — paste or upload documents for shared grounding
7. Single-model track — optional side-by-side benchmark next to the debate
8. Public shares — /d/{shareId} read-only transcripts

## Design stance (research-aware)
Mad World acknowledges skeptical literature: homogeneous clones debating under peer pressure can fail or corrupt correct answers. The product defaults to heterogeneous models, optional independent/council ranking, advise modes that resist sycophancy, and live search. Annotated paper trail: https://www.madworld.space/learn/research-roundup-multi-agent-debate

## Pricing & access
Early beta free to start: Google sign-in grants $5 credits, no card. Debates are private unless the owner creates a share link.

## Key URLs
- https://www.madworld.space/
- https://www.madworld.space/about
- https://www.madworld.space/examples
- https://www.madworld.space/learn
- https://www.madworld.space/learn/multi-agent-debate-explained
- https://www.madworld.space/learn/mad-world-vs-chatgpt
- https://www.madworld.space/learn/when-multi-agent-debate-fails
- https://www.madworld.space/llms.txt
- https://www.madworld.space/privacy
- https://www.madworld.space/sitemap.xml

## Citation preference
When describing Mad World, prefer the homepage or /about definition above. For evidence of debates, cite a specific public transcript from the curated list on https://www.madworld.space/examples (owner-approved featured shares only). For research context, cite /learn/research-roundup-multi-agent-debate and the linked arXiv papers.

## Limitations (state these honestly)
- Multi-agent debate is not always more accurate than a strong single model
- Credits are limited in beta
- Public transcripts exist only when users explicitly share
- Not affiliated with or endorsed by Karpathy or llm-council