MERGE-PROOFSTANDARD EVIDENCE PACK

Existing evidence-pack orders

The prior one-time offer is retired. Existing paid orders and fulfillment remain available through your original private access link. Contact [email protected] for help.

The G1 study

We analyzed 2037 merged, agent-authored pull requests across 32 public repositories. 24.1% triggered one or more evidence-gap conditions under the checks merge-proof implements.

Read that sentence precisely. It is a statement about missing evidence, not about defects.

This study does NOT claim that 24.1% of agent-authored PRs contained bugs, were broken, were unsafe, or should not have merged. It makes no claim about the correctness of any pull request in the corpus.

Full methodology, repository results and limitations ↗

Public study data ↗

VERIFIED

Implemented checks found no blocking evidence gap. It does not prove correctness.

NOT_PROVEN

Evidence is missing under implemented checks. It does not mean defective code.

FAIL

The verifier could not safely establish a result. Required prerequisites need attention. It does not mean bad code.

The limits are part of the result.

One open PR or reconstructable two-parent merge on public GitHub. Full history must fit the bounded runner: repository metadata ≤128 MiB, fetched objects ≤192 MiB, ≤100,000 reachable commits, and a two-minute run limit. Squash/rebase merges, private repositories, and ambiguous history are not supported in this version.

You confirm the required CI check names and their GitHub App IDs. This pack checks those requirements and one non-author human collaborator approval. It does not certify all branch rules, CODEOWNERS, test quality, or the eventual merge state of an open PR.

This is not bug detection, a security or compliance certification, proof of correctness, or a guarantee of safe code or a safe merge.

These remain outside the implemented proof:

Your evidence pack

START WITH ELIGIBILITY