We’re building Counterfactual Acceptance Testing: a developer tool that asks a simple question before you trust a green AI-written PR — would a plausible wrong version of the feature have passed the same tests? Given a human-approved business rule, we generate plausible alternative interpretations, find minimal cases where they disagree, and audit whether the current acceptance evidence can actually tell them apart. Instead of claiming software is “proven correct,” we surface surviving wrong interpretations as explicit evidence gaps, with replayable witnesses and held-out probes. Our first target is material TypeScript business-rule changes such as authorization, tenant boundaries, account states, entitlements and refunds. We’re building for developers who need more confidence in AI-generated code without adding another generic code reviewer.