← Build Week 2026 collection
Submitted Build Week prototype · Developer Tools

Release confidence becomes evidence you can inspect.

ProofLatch evaluates one bounded release packet against server-owned policy. Deterministic code decides BLOCKED or READY; GPT-5.6 can explain the result and prepare a small Codex repair brief without changing it.

StatusSubmitted on Devpost; public signed-out judge path and v1.0.0 source release available.
ProblemAgent-written code can look finished before required release evidence exists.
MechanismStrict packet → deterministic policy → explanation → bounded repair brief.
AuthorityOnly the deterministic evaluator may set BLOCKED or READY.
ProofLatch release decision desk showing failed evidence and a blocked verdict
Public-safe Atlas Checkout fixture showing a BLOCKED release. It is synthetic sample evidence and does not attest that a real command ran.
Problem and mechanism

A model can explain release evidence without becoming the release authority.

The evaluator checks required gates and source cleanliness first. Only then may the model translate those facts into plain-language risk and a Codex-ready repair brief tied to exact blocking check IDs.

  • InputVersioned policy plus a bounded evidence packet for one supplied commit.
  • DecisionPure deterministic assessment that remains valid if the model is unavailable.
  • ModelSchema-bound explanation with no verdict field and no permission to weaken a gate.
  • Re-checkChanged evidence or commit produces a new digest and a new receipt.
Verified public evidence

Inspect the same product from four directions.

The public judge path works signed out. The optional authenticated model path is separate and never changes the deterministic verdict.

What the work demonstrates

Truth, interpretation and repair are separate layers.

That separation makes the product useful when the AI path works and safe when it does not.

Policy

Server-owned gates

A client cannot rename, omit or make a required check optional simply to obtain READY.

Explanation

Bounded model value

GPT-5.6 turns evidence into a clear risk summary while every reference stays attached to a real blocker.

Repair

Evidence must change

Codex receives a focused brief, but a repair is not accepted until new evidence reopens the latch.

Authority boundary

The receipt is a change detector, not a truth certificate.

The SHA-256 digest binds the effective policy, supplied packet, evaluator version and assessment after capture. It is not a signature, trusted timestamp, source attestation or proof that a command actually ran. The demo fix is visibly a fixture swap.

Model mayExplain evidence and propose check-bound repair steps.
Model may notCreate, upgrade, downgrade or override the release verdict.
Build Week 2026

See the other three bounded-authority experiments.

Gatehold governs agent clearance, ExitCanary tests SaaS exit fidelity and AccessCrash tests whether a process remains humanly reachable.