About Council

Council is a multi-agent deliberation system that orchestrates 5 frontier LLMs to answer complex questions without confirmation bias.

The Problem

A single AI model has confirmation bias — it tends to reinforce assumptions in the prompt, agree with the user, and present a single coherent narrative that hides genuine uncertainty. When you ask Claude, GPT, or Gemini the same hard question, each gives you a confident answer that reflects its own training biases.

The Solution

Council queries 5 different model families simultaneously, has them anonymously peer-review each other, then synthesizes the strongest reasoning while preserving genuine disagreements. The result is an answer that surfaces tension rather than hiding it.

The 5 Agents

The Protocol

1. Clarify — Chairman refines the question with the user
2. Respond — All 5 agents answer independently in parallel
3. Review — Each agent peer-reviews the others (anonymized, no self-review)
4. Synthesize — Chairman combines best reasoning, preserves dissent, flags confidence

Anti-Bias Mechanisms

• Anonymized review: models evaluate responses without knowing who wrote them
• No self-review: agents cannot review their own responses
• SOUL.md: shared protocol instructing all agents to resist sycophancy
• Dissent preservation: chairman is explicitly told NOT to average opinions
• 5 model families: different training data = different blind spots

Research Basis

Based on 15 academic papers. See PAPER.md for the full literature review.

Source Code

Open source: github.com/danlex/council

127 automated tests. 42 bugs found and fixed across 4 review rounds.