Vesper: Evidence for AI Code Fixes

Developers reviewing AI-generated code face a practical challenge: knowing whether a reported bug is real and whether a proposed fix actually works. Checking requirements, reproducing failures and collecting test results takes effort. A “tests passed” message can also hide missing tests, skipped checks or changed assertions. Vesper connects these steps into one evidence-based workflow for developers and software reviewers. It brings together IBM Bob IDE, a local verification runner and an interactive React workspace. Users confirm a requirement, run the supported demonstration, inspect the finding and proposed patch, review execution records and record a separate human decision. The runner first checks the baseline tests. It then requires the original issue to reproduce twice, preserves the reproducing test unchanged during repair verification, and checks that the candidate passes both that test and the declared regression suite. The report connects the requirement, observed failure, exact code changes and test results. Vesper’s distinctive feature is the separation of AI proposals, execution evidence and human approval. It checks whether the expected tests actually ran and rejects incomplete or misleading evidence, including stale reports and skipped tests used to claim success. The working prototype uses a synthetic Java checkout project with a disclosed seeded defect and an existing candidate fix. A recorded run reproduced the original failure twice, passed the unchanged test against the candidate and passed all 12 regression tests. The verifier also produced the expected decisions in 16 synthetic verification scenarios. The current scope is one supported Java/Maven demonstration. It does not automatically discover and repair arbitrary repositories. Vesper aims to reduce investigation and review effort, but no productivity or accuracy advantage has been established. The next step is to measure human review effort on representative development tasks.