Contributing to OpenJev

Thanks for helping build an open System One decision engine.

Setup

uv venv --python 3.11 && source .venv/bin/activate
uv pip install -e ".[dev]" prompt_toolkit
pytest -q && ruff check src tests

Where to help

Look at the Roadmap. Good first issues:

• Multi-token label scoring (sum of log-probs) in backends/hf.py
• Shared-state prefill with KV-cache reuse across questions
• openjev serve (FastAPI, POST /v1/systemone)
• MLX backend for Apple Silicon
• Web Playground (side-by-side comparison)
• Eval harness on BoolQ / Banking77 with ECE

Rules of thumb

• Keep the request/response schema identical to the Jev docs so the official SDKs keep working.
• Every backend implements exactly one method: score_options(prompt, labels) -> (logits, n_input_tokens).
• No proprietary code, weights or data. This project reproduces an interface pattern, not TypeSafe's model.
• Be honest in benchmarks: report the model, revision, hardware and sample size.

Style

ruff format + ruff check. Type hints everywhere. Tests for anything that touches the schema.