Proof to Publish

Proof to Publish — Evidence-Dependent Approval for AI-Generated Advertising** An AI-generated advertisement can be correct when approved and wrong tomorrow. Prices change, delivery promises expire, but previously approved creatives can continue carrying outdated claims. Proof to Publish addresses this problem by making creative approval conditional on the underlying evidence. The application demonstrates this with NOVA One, a fictional portable speaker. Two advertisements begin approved: a promotional creative offering €89 and free next-day delivery, and an evergreen brand creative without offer-specific claims. When the merchant updates the offer to €109 and standard delivery, the system automatically holds the promotional creative while leaving the evergreen creative approved. It identifies the exact invalidated claims and the source records supporting them. Meta's Muse Spark then proposes a targeted correction, preserving the original Muse Image artwork and unaffected copy. The revised creative requires a separate human approval before it can be exported. The underlying technical contribution combines field-level source dependencies, versioned evidence, selective invalidation, constrained AI-assisted repair, and deterministic approval enforcement. The application is implemented in Next.js and TypeScript, with real Meta Model API integrations, executable tests, an inspectable event history, and verified export packages. It is a functioning standalone prototype, not a Meta Ads integration or automated publishing service.