What it does
Quorum runs a panel of small open-weight models on your machine. They review a diff blind, read each other's anonymized findings, and revise. A judge then rules, and the screen shows who changed their mind, who dissented, and whether the judge overrode the weighted vote.
Each model's vote is weighted by a per-category skill profile measured on a labeled corpus, using the Wilson 95% lower bound. A model can only change its mind by giving evidence. Anything in the pasted code that tries to steer the models is treated as data, and the input is flagged.
What's built
- A FastAPI backend with SSE streaming, calibration, and a held-out evaluation with confidence intervals across six baseline conditions
- Semgrep tool grounding, role routing, and an MCP server so coding agents can ask for a local second opinion
- A model manager for Ollama, OpenAI-compatible endpoints, and a DigitalOcean inference preset
- A React/TypeScript UI with a debate view, skill profiles page, and evaluation page
Open-weight models
It runs entirely on a laptop. We tested it with Qwen2.5 3B, Llama 3.2 3B and Gemma 3 4B through Ollama on an 8 GB Apple M2. The code is MIT licensed.
Built With
- fastapi
- ollama
- python
- react
- semgrep
- sqlite
- typescript
Log in or sign up for Devpost to join the conversation.