A traceable deliberation harness for AI-assisted research and technical decision-making.
DebateLab turns model disagreement, tool use, evidence gathering, critique, revision, and synthesis into inspectable system events.
It's not a chatbot. It's a research hearing where every participant has a role, every argument has evidence, and every conclusion can be inspected.
# Clone and start
git clone https://github.com/rmax-ai/debate-lab.git
cd debate-lab
docker compose up -d
# Open http://localhost:3000- You enter a topic, context, goal, and constraints
- The orchestrator selects specialized debate agents from a harness registry
- Agents research using read-only tools under policy governance
- Structured debate rounds with claims, evidence, cross-examination, and revision
- Final synthesis with auditable trace of every claim, tool call, and position change
User → Next.js/React UI → SSE/REST → FastAPI Backend
├── Debate Orchestrator
├── Agent Harness Registry
├── Tool Gateway
├── Evidence Extractor
├── Claim Tracker
├── Synthesis Engine
└── Eval/Audit Engine
│
Postgres (events, traces)
- Claims are first-class objects, not just transcript lines
- Every tool call is mediated through a policy-enforcing gateway
- Event sourcing makes every run replayable and auditable
- Mock-first development validates the protocol before real LLM costs
- Structured outputs — no free-text agent responses allowed
| Document | Purpose |
|---|---|
| ARCHITECTURE.md | Full system architecture |
| THREAT_MODEL.md | Failure modes and mitigations |
| ROADMAP.md | Phased delivery plan |
| DECISIONS.md | Key architectural decisions |
| AGENTS.md | Conventions for contributors and AI agents |
Frontend: Next.js, React, TypeScript, Tailwind CSS, shadcn/ui Backend: Python 3.12+, FastAPI, Pydantic v2, asyncio Storage: PostgreSQL (events, traces), S3-compatible (artifacts) Deployment: Docker Compose
v0.1.0 — MVP (in development) Mock-first core engine with full debate lifecycle, event sourcing, and UI.
MIT