Repair a formal proof by iterating on code and builds

Codex fixed CertiGC proofs in Rocq with human review on any change that could alter the specification.

Done with Codex· Advanced

What they did
A developer let Codex work inside the repository and edit Rocq proof files directly. It checked proof states through rocq-mcp, made targeted edits, and kept validating them with make builds. It handled the VST proofs first and then moved to the graph-isomorphism theorem. When it proposed a recorded-backward-edge invariant, the author reviewed that idea, and any Definition change required discussion before it could go in.
How it went
The branch compiled successfully and was accepted in PR 30. The main theorem was approved on April 29, and a later audit dropped a stale no-backward-edge premise. This was a single case, not a controlled study, and no speedup was measured.
Tips
This works best when a person stays in the loop and checks that any new invariant still matches the intended spec. Expect a lot of interaction if the proof tree is large.

Try it yourself with Codex

I have a [Coq/Lean/Isabelle] repository at [folder], and [file, lemma, or theorem] is failing after [change]. Open the proof state, make the smallest edits needed to get it working again, and rerun the relevant build after each change. If you believe a specification, invariant, or Definition must be adjusted, pause and ask me before changing it. Do not submit, merge, delete, or otherwise make an irreversible change until I explicitly approve it. At the end, give me a list of every file and statement you touched, plus any ideas you considered but did not apply.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on arxiv.org ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.