Mrdzyn Studio

**ProofPilot AI — Don't Just Fix the Bug. Prove It's Fixed.** AI coding agents can generate convincing bug fixes, but how do we know those fixes actually work? ProofPilot AI tackles this challenge by separating AI-generated proposals from independent, evidence-based verification. Using a synthetic e-commerce website, we demonstrate a mobile bug where a promotional overlay intercepts Add to Cart taps, leaving the cart counter unchanged while desktop behavior works correctly. ProofPilot reproduces the defect with a trusted Playwright browser oracle and binds the evidence to an exact Git commit. We integrated **Meta Muse Spark 1.3 through the Meta Model API** for native video understanding and structured fix proposals. After an incomplete initial analysis, Muse Spark successfully analyzed the recording and identified the observed symptom. A separately approved request generated a minimal CSS correction using a strictly constrained grammar. **Our innovation is the trust boundary:** the AI proposes, but deterministic verification decides. Model outputs are treated as untrusted data and cannot modify the trusted tests or award themselves a successful verdict. The system uses disposable verification stages, fresh browser replays, evidence hashing, negative controls, and strict authorization gates. A separately hand-authored correction passed **3 out of 3 independent browser replays**. The actual Muse Spark-generated proposal has not yet been executed or verified and is explicitly labelled **UNTRUSTED — NOT VERIFIED**. ProofPilot AI is currently a working hackathon prototype demonstrated on one synthetic fixture. Our vision is to make AI-assisted software development more reliable by replacing unverified claims with independently auditable evidence. **ProofPilot AI: The model proposes. Deterministic code decides.**