New to this work? Start here: https://github.com/LalaSkye/start-here
Reference kernel for execution-boundary governance.
Commit Gate Core stops unauthorised consequences before execution.
Most governance systems review decisions after consequences happen.
This repo demonstrates a smaller, harder control surface:
No state mutation is allowed unless a signed, scoped, unexpired, unreplayed
DecisionRecordauthorises the exact commit.
If authority, scope, expiry, replay, or receipt checks fail, the action does not run.
The proof is the failed execution path.
This repository demonstrates a path-local commit gate.
It enforces a single invariant at the boundary it sits on:
No consequence at this commit boundary without a valid signed, scoped, unexpired, unreplayed DecisionRecord — payload-bound, proof-bound, atomic commit.
This is the path-local invariant. It is what the code in this repository implements and tests.
The path-universal invariant is stronger:
No consequence is reachable without passing a proof-bound, payload-bound, atomic commit boundary across all paths.
Achieving the path-universal invariant is an architectural placement question, not a gate-implementation claim. It requires:
- routing every reachable path to a consequence through a gate of this kind
- exclusion or explicit out-of-scope marking of alternate routes, including:
- human review handoffs
- downstream agent execution
- asynchronous side channels
- bypass paths created by retries, rollbacks, or recovery flows
- system-level evidence that the routing holds
This repository does not make the path-universal claim.
- If you want to inspect the gate primitive: read
cmd/,internal/gate/, and the test suite. - If you want to evaluate path-universal coverage in a real system: that is a deployment-architecture review, not a code review of this repo.
Use the test to check whether a system can physically stop consequence at the point an action would become real.
See: docs/execution-boundary-test-v1.md
Core question:
Where does the system physically stop?
PASS: The action cannot execute without valid proof.
FAIL: The action still reaches consequence.
git clone https://github.com/LalaSkye/commit-gate-core.git
cd commit-gate-core
python -m examples.unsafe_email_sendExpected output:
Result: HOLD
Email sent: false
Receipt written: true
If the email sends, the gate is broken.
Attempt: send external email
DecisionRecord: missing authority
Result: HOLD
Email sent: false
Receipt written: true
That is the shape.
The system refuses the unsafe state change before execution and writes a receipt proving why.
- Unsafe consequence can be refused before execution.
- Missing authority prevents mutation.
- Refusal can produce an auditable receipt.
- Bypass failure can be tested directly.
This is not governance commentary.
It is a small enforcement primitive.
This repo does not claim to be a full AI governance system.
It proves one narrow invariant:
This path cannot execute without a valid
DecisionRecord.
The invariant is deliberately small so it can be inspected, tested, and broken under hostile reading.
This repository does not prove adoption, certification, standardisation, or production readiness.
It demonstrates a bounded execution-control surface that can be run, inspected, and tested.
A valid DecisionRecord must be:
- signed
- scoped to the exact commit
- within its validity window
- unreplayed
- sufficient for the requested mutation
Failure at any check produces HOLD.
No silent continuation.
A useful governance gate must show:
- what action was attempted
- what proof was required
- which check failed
- whether execution occurred
- what receipt was written
For this demo, the answer is simple:
Execution occurred: false
Receipt written: true
Verdict: HOLD
v0.1 — one invariant, hard-enforced.
Small surface. Clear failure mode. Receipts over reassurance.
From Policy to Commit: Execution-Boundary Control for Governed AI Systems
- DOI: https://doi.org/10.5281/zenodo.19980275
- Zenodo record: https://zenodo.org/records/19980275
- PDF (in this repo): docs/papers/From_Policy_to_Commit_Ricky_Dean_Jones_AlvianTech_Working_Paper_v0.1.pdf
Jones, R. D. (2026). From Policy to Commit: Execution-Boundary Control for Governed AI Systems (v0.1). Zenodo. https://doi.org/10.5281/zenodo.19980275
MIT. Use it. Break it. Tell me how.