PRoove

AI code review has a trust problem: a reviewer can confidently flag a bug without proving that the pull request introduced it. Developers then have to reproduce the issue, compare revisions, and test the suggested repair themselves. PRoove brings that work into the GitHub pull-request workflow. When a PR opens, an AI model proposes one hypothesis and one regression test grounded in the project's documented behavior. PRoove runs the same test against the PR's actual base and head commits in separate Git worktrees. It reports PROVEN BUG only when the test passes before the change and fails after it. Test results, not the model's confidence, determine the verdict. Our working checkout demo makes the difference visible. A one-line PR removes the lower bound from a discount calculation. A $100 item with a 150% discount now returns −$50, although the specification requires a non-negative price. PRoove reproduces the failure and posts a Proof Card in the PR with the commit SHAs, test results, specification rule, observed error, logs, and a link to the GitHub Actions run. PRoove then tests a proposed one-line repair on a temporary copy of the changed code. Only a passing repair is offered as a native GitHub suggestion. A developer reviews and applies it; PRoove reruns the original regression test and updates the card to FIX VERIFIED when that specific issue is resolved. The full cycle has been demonstrated in real GitHub pull requests. PRoove is for developers and teams that want AI-assisted reviews they can inspect and reproduce. Its distinctive approach is to make executable evidence the gate between an AI hypothesis and a human-approved fix. The current prototype supports a controlled TypeScript/Vitest checkout project, same-repository PRs, and single-line suggestions. We plan to extend its proof workflow to more repositories and test environments. The model proposes. Tests prove. A human approves.