The brief
What it
does.
- Gives each task an explicit question, named owners and reviewers, decision authority, evidence references, and a visible disagreement path.
- Uses a provider-neutral protocol and Git-backed records so work can move across models, vendors, local tools, and human review without losing provenance.
- Treats model output as untrusted input: Relay coordinates work, but it does not decide whether a scientific claim is true or replace peer review.
Retrospective
Still
becoming.
Relay grew out of the practical problem of fragmented AI work: too many chats, copied prompts, and decisions that were hard to reconstruct later. The project keeps becoming more useful when it makes authority, evidence, and uncertainty more visible instead of hiding them behind a smoother interface.