Compare four states
Base, change A, change B, and their combination. Keep revisions and declared checks identifiable.
Two changes can pass their tests separately and still break a requirement together. MergeWitness investigates that interaction and records a counterexample you can replay.
A demonstrated synthetic fixture: tenant-specific pricing meets a cache keyed only by SKU. Alpha's cached price can reach Beta.
A tenant must receive its own price, regardless of which tenant made the previous request.
The value is the sequence: Alpha first, then Beta, for the same SKU. Ordinary checks in the fixture remain green while the explicit behavior probe exposes the combined defect.
The verifier evaluates observed results. An AI assistant may help propose a repair; it does not decide that the repair passed.
Base, change A, change B, and their combination. Keep revisions and declared checks identifiable.
Run an explicit behavioral check. Record what the probe exercised, expected, and observed.
Keep a reproducible sequence and its evidence. A result is meaningful within the probe's scope.
Verify against frozen declared checks and feature checks. A candidate passing those checks is a bounded result.
MergeWitness is a working open-source prototype created for the IBM Bob 2.0 Hackathon. Its original submission is preserved while judging is unresolved.
Read the frozen prototype documentation ↗Node.js core, CLI, MCP server, React lab, tests, retained evidence, two synthetic scenarios, and an MIT license.
Immutable repair attempts, broader evidence bundles, sandbox validation, developer experience, and optional Codex, Claude, and Gemini adapters.
External repository execution requires isolation work. No witness found by a probe does not certify a merge as safe. Product adoption and comparative performance remain to be established.
Start with the documented synthetic cases. Inspect the reports, understand the limits, and help shape a tool whose conclusions come with evidence.
Explore the open-source prototype ↗