A generic AI check looked at the wrong red and rated it 0.97 on-brand. It would have shipped. Real color science caught it at the door. We built the Brand Fidelity Firewall, a pre-ship gate every AI-generated campaign asset has to clear before it goes out. One asset in our test batch used a red that looked close enough, and a common image-similarity check (a perceptual hash, the "looks-fine" score much AI tooling leans on) rated it 0.968: a 0.97 "on-brand" pass. Except it was the wrong red. In real CIEDE2000 (ΔE00, the ISO/CIE-standard for color difference), it sat 7.12 off the brand's Lumiere Crimson, Pantone 484, against a tolerance of 3.0. A similarity guess cannot tell your exact red from a competitor's. Color science can, so the gate blocked it. The decision is deterministic code outside the model. The AI advises, the gate measures and returns PASS, FLAG, or BLOCK, each backed by a reproducible number. (The "Lumiere" brand and these assets are synthetic, built so every check measures real pixels.) Before an AI-made asset ships, does anything measure it against your exact Pantone spec, or does a "looks on-brand" score decide? We'd genuinely like to hear how your team handles it. #BrandGovernance #CreativeOps #Pantone #MarketingCompliance #BrandSafety