PromiseProof
AI can fix code. It shouldn't certify itself.
Start new thread

PromiseProof - AI can fix code. It shouldn't certify itself.

by•
PromiseProof shows a safer way to use coding agents. GPT-5.6 investigates a broken product promise, Codex proposes a constrained repair and a human approves the exact patch. Neither model can award PASS, the unchanged deterministic verifier does. Try it in your browser, tamper with the evidence, reproduce a sealed report or run the same gate in CI.

Add a comment

Replies

Best
Maker
📌

Hey Product Hunt 👋

I built PromiseProof during OpenAI Build Week because I kept returning to one uncomfortable question: when an AI agent repairs code, who decides that the repair is actually correct?

In PromiseProof, GPT-5.6 investigates a sanitized evidence dossier and selects one bounded diagnostic replay. Codex proposes a constrained source patch in isolation. I review and approve the exact diff. Then the same unchanged deterministic test that caught the broken promise decides PASS or FAIL and neither model can grade its own work.

The current release proves the workflow on an 'off means off' personalization contract, with a browser verifier, CLI, and GitHub Action.

I would especially appreciate feedback from people building coding agents or working in QA, privacy, reliability and CI.

OpenAI
OpenAI Day
What could you build with GPT-5.6 that wasn’t practical before?
Maker
GPT-5.6 made it practical for me to build and ship an end-to-end, auditable AI repair system in a single Build Week. PromiseProof can reason across the UI, browser storage, network traffic, backend state and source code to diagnose failures that fixed rules cannot reliably untangle. GPT-5.6 analyzes a sanitized evidence dossier, ranks competing root causes, and selects one tightly controlled diagnostic replay through strict function calling. Codex then proposes a constrained patch, a human approves the exact change, and an unchanged deterministic verifier decides PASS. The result is more than AI that writes a fix: it is a complete repair workflow teams can inspect, reproduce, and trust without ever letting the model certify its own work.