Close Loop Protocol

A governance protocol for running AI builders on real repositories — and a fully archived case of it working.

AI writes the code now. The scarce thing is no longer code volume — it is knowing whether the output is true. This repo contains a protocol built and stress-tested on a live multi-repository system, where AI executors implement and every claim is bound to verifiable evidence.

Core mechanisms

• Bounded slices — every task ships with an explicit maximum completion claim and a forbidden list
• Read-back gate — no code before the executor restates the task as a machine-checkable tuple
• Exact-head evidence — tests and CI bind to exact commit SHAs; "passed" must say what, on which revision, passing which checks
• Fail-closed reporting — BLOCKED is a first-class, honored outcome, never a failure to be hidden
• Human-only merge — AI recommends; humans merge

What's inside

• protocol/SPEC.md — the protocol: roles, permission layers, state machine, evidence rules
• cases/case-574.md — the complete archived record of one slice: three substance-level drifts (including an all-green test suite built on fabricated fixtures), three catches, three honest BLOCKEDs, and the real baseline capability gap the run uncovered
• templates/ — copy-ready formats: checkpoint, read-back, approval, blocked report, merge receipt

Who this is for

Teams putting AI coding agents into production who need to know — not hope — that the output is what it claims to be.

License

MIT — see LICENSE.