TypeSafeAI/jev-harness ? reverse-engineered prompt

Build me a TypeScript project that acts like a proposal review harness for an AI coding agent.

I want the model to make one proposed action, then a separate Jev style step answers a few narrow questions, and the code decides whether the proposal should go forward. It should stay safe and opinionated, with no automatic patching, no running proposed code, and no hidden permissions. Include the core contract, validation, deterministic decision logic, and a receipt or evidence trail for every review.

Also add an offline demo so I can try the flow locally in a browser, plus a small benchmark or fixture runner that compares proposal review results against synthetic examples. Keep it able to run without an API key for the offline checks, and make the live provider pieces clearly host controlled. If you need current docs or best practices for anything, look them up online first.