- A generic similarity score looked at an off-brand campaign asset and rated it 0.97 on-brand. It would have shipped. Real color science measured the same red 7.12 ΔE00 off the brand's Pantone spec and blocked it. Same pixels. Opposite verdicts. 🧵
- The asset was AI-generated holiday creative for a synthetic luxury brand we built to test this, "Lumiere." Its hero red filled 45.9% of the frame at hex 9D2E0B. Close enough to fool a similarity guess. Wrong enough to be off-brand.
- The brand book pins exactly one red: Lumiere Crimson, PMS 484, hex 9E2B25, with a per-color tolerance of 3.0 ΔE00. Anything past that line is not the brand's red. The gate does not negotiate on it.
- ΔE00 is CIEDE2000, the ISO and CIE standard for perceptual color difference. Not an aesthetic vibe. Our implementation is validated against the Sharma et al. reference pairs, part of a 5/5 passing unit suite.
- The generic baseline we run is a perceptual hash, DCT based, driven by luminance and structure. It cannot tell a correct Pantone red from a competitor's. We put it beside real color science so the failure mode is measurable, not theoretical.
- So on that asset: the perceptual hash says 0.968, on-brand, ship it. ΔE00 says 7.12 against a 3.0 tolerance, off-brand, BLOCK. One of those numbers would have waved a wrong red straight into a holiday campaign.
- The verdict is deterministic code sitting outside the LLM. The model advises on tone and abstains when unsure. It never changes the result. PASS, FLAG, or BLOCK, each backed by an objective measurement, not a guess.
- This is why it holds at any model quality. A flawless generator still ships assets that must be measured against exact PMS specs. Better models just make prettier wrong reds. The gate measures either way.
- On our fixed 12-asset demo batch: 7 auto-cleared, 3 blocked, 2 flagged to a human. Every cleared asset exports a Brand Fidelity Certificate (JSON and printable HTML) with the per-check measurements. A receipt, not a rubber stamp.
- For the brand-ops and AI-builder crowd: when a generic similarity score and a color-science gate disagree on whether an asset is on-brand, which one should hold the authority to block the ship? #AIGovernance #BrandCompliance #ColorScience #CreativeOps
- We built the whole gate to be inspected, not trusted on faith. The demo scores the batch live and opens the evidence panel behind every verdict: https://veriprajna.com/demos/brand-ai-content
Published on X · July 22, 2026
On social media
See this post on its original platform
In our archive