Repair formal proofs for a verified garbage collector
Codex edited files and ran builds to repair CertiGC proofs; it handled local repair well, but humans had to check that its invariants preserved the specification. Task: Suggested implementation: Use Codex to edit Rocq proof files and run targeted make builds. The agent should inspect proof states via a rocq-mcp server, prioritizing VST proof repairs before graph-isomorphism theorems. Human reviewers must approve all Definition changes and verify that proposed invariants preserve the specification.; tools: Codex. Human role: Source evidence is unavailable for specific human role definitions; however, the input indicates humans were required to verify that invariants preserved the specification, review proposed invariants, and enforce strict change controls.; limitations: Prerequisites include repository access, Rocq environment, and build tools. Expect heavy supervision with high tool call volumes. Human approval is mandatory for any invariant changes or new preconditions. Source reliability is low due to blocked access, so treat all metrics as estimates..
Tools
Published on Agent Work Atlas:
Overview
Codex edited files and ran builds to repair CertiGC proofs; it handled local repair well, but humans had to check that its invariants preserved the specification.
Workflow
Suggested implementation: Use Codex to edit Rocq proof files and run targeted make builds. The agent should inspect proof states via a rocq-mcp server, prioritizing VST proof repairs before graph-isomorphism theorems. Human reviewers must approve all Definition changes and verify that proposed invariants preserve the specification.
Reported result
Actual outcomes are unknown as the source is blocked. A measurable acceptance check is required: confirm the branch compiles successfully and merges upstream. Track the ratio of active agent hours to human prompts to assess supervision overhead, as no speedup was previously measured.
Human role
Source evidence is unavailable for specific human role definitions; however, the input indicates humans were required to verify that invariants preserved the specification, review proposed invariants, and enforce strict change controls.
Limitations
Prerequisites include repository access, Rocq environment, and build tools. Expect heavy supervision with high tool call volumes. Human approval is mandatory for any invariant changes or new preconditions. Source reliability is low due to blocked access, so treat all metrics as estimates.
Try it
Act as a formal verification engineer. I need you to repair proofs in a verified garbage collector repository. [Replaceable inputs: Repo URL, specific Rocq files, build commands]. First, perform a dry run on a small subset of proofs to verify the rocq-mcp connection and build process. Do not modify any Definition files without my explicit approval. Propose invariants for the graph-isomorphism theorem and wait for my review. If the build fails, report the error log immediately. Only proceed to full repair after I confirm the dry run results. Ensure you track the number of tool calls and human interactions for overhead analysis.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.