Differential Oracle

For a vulnerability instance, two builds of the same software are prepared: vulnerable (b⁻) and patched (b⁺). A candidate input x is executed on both builds under sanitizer instrumentation. The oracle accepts x as a valid PoC only if b⁻ triggers the target sanitizer failure and b⁺ does not. Because the predicate is programmatic and decidable, an agent can test each candidate without supervision; the verdict together with the sanitizer trace provides grounded feedback, turning PoC construction into an executable propose–verify–refine loop.