RealDiff finds runtime behavior changes that ordinary source review misses. It builds two Git revisions, observes their tests, learns a noise baseline from three base runs, and reports the first changed behavior in each call tree. A source diff tells you what was edited. RealDiff tells you what the edit did, including effects in files the pull request never touched.

A harmless-looking refactor can change behavior far from the edited file. For example, replacing List.Sort with OrderBy alters an unedited pricing engine. In the demo, Discount Engine.Select Discount returned "CLEARANCE_40", now returns "SEASONAL_15". Checkout Totals.Compute returned 60, now returns 85. The edited helper is in Infrastructure.Collections; the observed effect is in Commerce.Pricing. Two of the three tests that executed this did not assert on the change.

Run the included demo or inspect the maintained public .NET demo pull request and its successful hosted run. The proof creates a temporary proposed-change tree, changes only Sorting Extensions.cs, runs the base twice plus the change once, and writes findings.json. It verifies the edited file contributes zero traced members, the frontier is Commerce.Pricing.Discount Engine.Select Discount in an unedited project, two call sites changed without an assertion reacting, five diverged keys collapse to three frontier nodes, and equal-priority selection is deterministic across fresh processes.

The maintained .NET, Java, Node, Go, Rust, and Python gates apply the same conformance rules: identical method sets, per-key event counts and entry ordinals, source tripwires, digest proofs, and zero engine divergences from non-empty runs. Status: early preview.