⚖️ Tribunal

Adversarial multi-agent fact-checker with just-in-time RAG grounding.

Give it a factual claim → get a verdict (True / False / Misleading / Unverifiable) with a confidence score and citations to real sources. Built to showcase the 2026-in-demand stack: LangGraph multi-agent orchestration, RAG, an MCP server, Langfuse observability, and evals.

Pipeline (v1)

claim → Decomposer → Retriever (RAG) → Analyst → Judge → verdict

• Decomposer — splits the claim into atomic sub-claims + search queries
• Retriever (RAG) — Wikipedia-first hybrid search → chunk → MiniLM embed → LanceDB → top-k
• Analyst — weighs supporting vs refuting evidence, using retrieved sources only
• Judge — final verdict + confidence + verbatim citations (structured via Instructor)

Stack

LangGraph · OpenRouter (free models) · Instructor · fastembed MiniLM/ONNX (local, ~50MB) · LanceDB · Langfuse · FastMCP · FastAPI. No local LLM, no torch — runs on an M1/8GB.

Setup

cd tribunal
uv venv && source .venv/bin/activate
uv pip install -e .          # add -e '.[evals,dev]' for RAGAS + pytest
cp .env.example .env         # then paste your OPENROUTER_API_KEY

Run — three interfaces

# 1. CLI
tribunal "the Great Wall of China is visible from space"

# 2. MCP server (wire into Claude Desktop / Cursor as an MCP tool `verify_claim`)
tribunal-mcp

# 3. Web UI
uvicorn tribunal.web:app --reload      # http://127.0.0.1:8000

Evals

python evals/run_evals.py     # verdict accuracy vs evals/golden.jsonl

Observability

Set LANGFUSE_PUBLIC_KEY / LANGFUSE_SECRET_KEY in .env and every run streams a full hierarchical trace (per-agent spans, tokens, cost, latency, retrieved chunks) to Langfuse. Leave the keys blank to disable — the pipeline runs identically without it.