Portfolio
LRPROOFLOOP // AGENT REVIEW LAB

CODEX SUBSCRIPTION READY

OPENAI BUILD WEEK / DEVELOPER TOOLS

Make agent-written code earn acceptance.

ProofLoop turns a code diff or failure report into independent GPT‑5.6 review in Codex, a bounded correction plan, regression evidence, and a final decision that only a human can make—without an API key.

Run a review Inspect the control loop
01

SUPPLIED EVIDENCE

Choose a failure boundary

CRITICALDOMAIN CORRECTNESS

Payment response misclassification

A gateway returns HTTP 200 for a declined payment. The first implementation treats transport success as payment success and advances the transaction state.

Acceptance criterion
Only a confirmed provider approval may transition a payment to completed.
First-pass gap
No mapping between provider result codes and the domain state machine.
review-input.txtEDITABLE

0 / 12,000

Edit the sanitized sample or paste your own material. Do not include secrets, credentials, or private customer data.

02

CONTROL LOOP

Run independent review

  1. 1BoundGoal + evidence
  2. 2ReviewCodex panel
  3. 3CorrectFindings → plan
  4. 4GateRegression evidence
  5. 5DecideHuman sign-off
evaluation.log

READY Evidence loaded. Copy the prompt into Codex or run the judge demo.

The repo-scoped $proofloop-review skill uses your Codex subscription; no API key is required.

WAITING0 / 5 gates complete

MODE / NOT STARTED
03

JUDGE PANEL

Traceable verdicts

DOM

Domain reviewer

State and business rules
PENDING

Waiting for an evaluation run.

Evidence
—
Next action
—
SEC

Security reviewer

Trust and isolation
PENDING

Waiting for an evaluation run.

Evidence
—
Next action
—
UX

Product reviewer

User-visible behavior
PENDING

Waiting for an evaluation run.

Evidence
—
Next action
—
TST

Test reviewer

Coverage and regression
PENDING

Waiting for an evaluation run.

Evidence
—
Next action
—
04

ACCEPTANCE EVIDENCE

Prove the correction

BOUNDED CORRECTION

Run a review to reveal the correction plan.

ProofLoop will separate the proposed change from evidence that still requires verification.

  • No changes proposed yet.

REGRESSION PLAN

  • Evidence is generated after review.

HUMAN CHECKS

  • The final gate remains a human decision.

GATE 05 / HUMAN AUTHORITY

No automated acceptance.

After reviewing the findings and evidence plan, a human must accept or reject the proposed disposition.

PENDINGHuman decision required

WHY THIS IS A DEVELOPER TOOL

Reliable AI engineering is a control system.

GPT‑5.6 in Codex provides judgment. The repo-scoped $proofloop-review skill runs four evidence-labeled review gates through a ChatGPT subscription—no API key or runtime API charge required.

ProofLoop provides constraints. It limits supplied context, refuses to invent executed tests, makes evidence gaps visible, and preserves the human acceptance boundary.

Codex provides the build trace. The working product, tests, documentation, and submission materials were completed as one bounded Build Week implementation.

Inspect the source repository