mutual-dissent 0.2.0

Cross-vendor multi-model debate and consensus engine for AI response distillation

Mutual Dissent

Cross-vendor multi-model debate and consensus engine for AI response distillation.

NiceGUI web interface with live debate view, research dashboard, and full CLI research tool. Direct vendor APIs, replay, ground-truth scoring, cost tracking, and markdown export. 400+ tests across Windows and Linux CI.

Research program by Richard Spicer · GitHub

---

Install

pip install mutual-dissent

Or from source:

git clone https://github.com/q-uestionable-AI/mutual-dissent.git
cd mutual-dissent
uv sync --group dev

---

How It Works

1. Fan out — Query goes to Claude, GPT, Gemini, and Grok (direct APIs or via OpenRouter)
2. Reflect — Each model sees the others' responses and argues back
3. Synthesize — A user-selected model distills the debate into a final answer
4. Log — Full debate transcript saved as structured JSON with cost and token data

Why Cross-Vendor?

Single-vendor multi-agent systems (Grok's 4-agent debate, Anthropic's agent teams) share the same training data and blind spots. Cross-vendor debate surfaces disagreements that correlated architectures can't — different training data, different safety postures, different failure modes.

Usage

# Run a debate
dissent ask "Your query here"

# With explicit panel and options
dissent ask "Your query here" --synthesizer claude --rounds 2 --panel claude,gpt,gemini
dissent ask "Summarize this" --file report.pdf
dissent replay <transcript-id> --synthesizer grok
dissent serve
dissent config test

mutual-dissent also works as the full command name. Full documentation at docs.mutual-dissent.dev.

---

Status

Capability	Status
Core debate loop	✅ Complete — fan-out, reflection, synthesis via OpenRouter
Provider abstraction	✅ Complete — direct Anthropic API, mixed-panel routing
CLI research tools	✅ Complete — replay, scoring, cost tracking, markdown export
Web GUI	✅ Complete — NiceGUI debate view, research dashboard, live streaming
Documentation	✅ Complete — Mintlify docs site with AI assistant and MCP server
Desktop app & batch mode	Planned — Tauri wrapper, alternative topologies, public release

Research Platform

Full debate transcripts are logged as structured JSON for research — disagreement patterns, convergence dynamics, consensus poisoning, and hallucination detection. See Research Methodology for details.

Legal

All tools are intended for authorized security testing only. Only test systems you own, control, or have explicit permission to test. Responsible disclosure for all vulnerabilities discovered.

License

MIT

AI Disclosure

This project uses a human-led, AI-augmented workflow. See AI-STATEMENT.md.