Proof repair changes the theorem? Check assumptions first

Use Codex with a Rocq test project and proof obligation to prepare a minimal proof repair and checker logs through the test workflow below, then review it. Matching tools and permissions are required.

Execution environment and usage

AI / software

Codex

Where to operate

Codex CLI in a terminal, opened in a separate test folder.

Getting started

  1. Follow the linked official Codex CLI setup guide; open a terminal in the test folder, run codex and sign in.
  2. Supply a Rocq test project and proof obligation and the prompt below. Confirm the scope before allowing necessary actions.
  3. Inspect the resulting minimal proof repair and checker logs; record anything not run or not verified.

Setup and plan requirements

Requires access to Codex, authorized test data and a matching runtime; file writes and shell commands must stay within approved scope. Preparation: An expert reviewer, reproducible proof project and version-pinned test branch are prerequisites. A pinned Rocq toolchain, project dependencies and a formal-proof reviewer are required. Ordinary chat cannot replace kernel checking.

Prompt portability limits

Adapt the wording to another coding assistant only after separately configuring files, runtime, test tools and integrations. Renaming the tool does not establish compatibility.

Usage limitations

This workflow was not hands-on tested. Work on test copies; no automatic publishing, payment or production-data changes. Separate plans, actual results and untested areas.

Environment sources

適用情境 / When to use

Use Codex with a Rocq test project and proof obligation to prepare a minimal proof repair and checker logs through the test workflow below, then review it. Matching tools and permissions are required. Shengyi Wang reports using Codex to help complete and stabilize Rocq proofs in CertiGC.

步驟 / Steps

  1. An expert reviewer, reproducible proof project and version-pinned test branch are prerequisites.
  2. Select an unfinished obligation and explain its definitions and assumptions before proposing a repair.
  3. Rerun the kernel and compare theorem statements, premises and axioms line by line.

提示詞 / Prompt

Address [proof obligation] in [test branch] while preserving [theorem and allowed assumptions]. List dependencies, propose a minimal repair and supply kernel-check logs. Stop and explain if specifications or axioms must change. Return a diff for expert review.

留意事項 / Cautions

  • New axioms or skipped proofs do not count as completion.
  • Do not infer whole-system safety or deployment readiness.

來源限制 / Evidence limits

This experience report is not a controlled experiment; kernel acceptance does not show that a specification captures real needs.

流程說明 / Workflow note

These steps and prompts are original, untested exercises by this site, not the author’s verbatim workflow. Reported results do not guarantee exercise outcomes.

原始工具組合 / Original tool stack

Codex; Rocq / Coq; CertiGC; VST / CertiGraph

Source publication date: 2026-06-19

Source checked: 2026-10-19

AI-Assisted Completion of CertiGC Proofs: An Experience Report