RealDiff diffs runtime behavior instead of source, across six languages.

It checks out both branches into isolated worktrees, builds each with instrumentation, runs the same tests on both sides, compares actual method arguments and return values, filters noise by running the base repeatedly, then collapses results to the frontier where a change originates. Its documented example follows execution from an edited ordering helper into unedited pricing code and reports the changed values there, which is the class of regression a source diff will never show. .NET via Mono.Cecil, Java via a javaagent with ASM, Node/TypeScript via Babel, Go and Rust via AST rewriting, Python 3.12+ via PEP 669. It only asks an LLM to explain findings it already proved deterministically, which is the correct division of labor. (GitHub)