Code Tribunal

An intent-conformance court for AI-generated code — in the browser, in your CI, and in your coding agent.

What this is

Tribunal puts AI-generated diffs on trial: seven coordinated agents compare the ticket against the implementation, catch omissions and scope drift, and issue a merge verdict with a 0–100 trust score and a traceability ledger. It runs in the browser as the War Room, and ships as a CLI and MCP server you can wire into Claude Code, Codex, or Cursor to gate real pull requests.

Why this exists

Linters check style. Tests check behavior. Nothing checks whether the diff actually does what the ticket asked — and that is exactly where AI-written code fails: it produces plausible, working code that silently skips a requirement, or changes something nobody authorized. Tribunal holds the ticket and the diff side by side and prices the difference. The verdict is deterministic arithmetic over agent-gathered evidence, so the same diff always gets the same answer.

Code Council

The Council editor is the companion sandbox: instead of one AI verdict, it streams several frontier model opinions over the same code in parallel and shows where they agree, disagree, or miss things entirely — with consensus signals, static scans, and multimodal input.

Stack

· Next.js 15 + Tailwind
· FastAPI + SSE streaming
· CLI + MCP server (uvx)
· Groq, Cerebras, Gemini · coordinated over Band