Model Council

Model Council is a debate-style app where GPT, Grok, and Claude discuss a question, critique each other, and produce a final verdict using judge-rubric scoring plus multi-domain verification.

Features

• Three-way model debate (GPT, Grok, Claude) using provider plugins/adapters
• Council modes:
  • Fast (single round)
  • Deep (three rounds)
  • Adversarial (one model plays red-team)
  • Research (source-grounded mode with citation requirement)
• Judge-rubric scoring (correctness, reasoning quality, evidence quality, robustness)
• Multi-domain verifier pipeline:
  • deterministic math checks
  • citation reachability checks
  • code-block sandbox checks (JS/TS)
  • logic consistency checks for contradictions
• Disagreement map with pairwise conflict/alignment signals
• Streaming debate feed (/api/council/stream) for live round updates
• Run history and benchmark analytics (/api/history, /api/benchmarks)
• UI to enable/disable models, choose model IDs, and add API keys
• API keys stored in browser localStorage only

Run

npm start

Then open http://localhost:3000.

API Endpoints

• GET /api/providers
• GET /api/modes
• POST /api/council
• POST /api/council/stream (SSE)
• GET /api/history
• GET /api/benchmarks

API Keys

Add keys in the UI for:

• OpenAI (GPT)
• xAI (Grok)
• Anthropic (Claude)

Keys are sent only when you run a debate and are not saved on the server.

Notes

• History is persisted to council-history.json in the project root.
• Verifiers are deterministic heuristics; they improve trust but are not formal proof systems.
• You can customize model names in the participant cards.