We scored 12 AI/ML platforms against a design contract. The infrastructure beats the models.

#webdev #design #a11y #ai

After scoring 30 websites, 16 dev tools, and 11 fintech sites against our 40-check design contract, we turned to the AI/ML sector. These are companies built by engineers who care about developer experience. They should score well.

They don't.

The leaderboard: 12 AI/ML platforms

Rank	Site	Score	Grade	Pass	Fail
1	Vercel AI SDK	73.9%	D	16	5
2	Vercel	67.4%	D	14	6
3	Modal	63.0%	D	13	7
4	Replicate	60.4%	D	12	7
5	Anthropic	60.0%	D	12	7
6	LlamaIndex	58.3%	F	11	7
7	Mistral	56.5%	F	11	8
8	Hugging Face	56.2%	F	10	7
9	Cohere	56.0%	F	11	8
10	Together AI	52.1%	F	11	10
11	Pinecone	46.0%	F	8	10
—	Perplexity	BLOCKED	—	—	—
—	OpenAI	BLOCKED	—	—	—

Zero A-grades. Zero B-grades. Zero C-grades. The best AI/ML platform scored 73.9% — a D.

The infrastructure-models split

The top 4 are all infrastructure: Vercel AI SDK, Vercel, Modal, Replicate. The bottom 6 are model companies: Cohere, Together AI, Pinecone, Hugging Face, Mistral, LlamaIndex.

This isn't coincidence. Infrastructure companies build developer tools — dashboards, docs, deployment UIs. They invest in design systems because their product is the interface. Model companies compete on benchmarks and API latency. The website is marketing, not product.

The scores reflect that priority gap.

What the leaders got right

Vercel AI SDK (73.9%) is the highest-scoring site we've seen outside of Designesy itself. Full Cadence typography — text-wrap: balance + pretty, tabular-nums, font-synthesis: none, line-height by role. Strong poise and takt: fine-pointer hover guards, press settle scales, concentric border-radius sets. 37 :focus-visible rules. Its main failures are will-change containing height and px-based font sizes.

Vercel (67.4%) shares the same design DNA — 158 :focus-visible rules, full Cadence, font-synthesis: none. It loses points on 3 transition:all instances and will-change containing width/height.

Modal (63.0%) is clean — no transition:all, will-change restricted to transform/opacity, full Cadence, reduced-motion support. It fails on no ::selection rule and no tabular-nums.

Replicate (60.4%) has full Cadence, text-wrap: balance+pretty, reduced-motion, no transition:all, clean will-change. But it fails hard on accessibility: button contrast is 1.00:1 — the button text is the same color as the button fill.

What the model companies got wrong

Together AI (52.1%) has press scales of 0.66 and 0.33. The contract floor is 0.95. When you press a Together AI button, it compresses to a third of its size. That's not a press interaction — that's a glitch.

Cohere (56.0%) has --ink luminance at 165/255. The "dark text" token is medium-gray. Button contrast is 1.00:1 — same as Replicate. But Cohere also has --ink on --paper at 2.54:1, failing WCAG AA across the board. The token system is misconfigured at the root.

Pinecone (46.0%) has no :focus-visible rules at all. Keyboard users cannot see where they are on the page. Its --paper token is #fff but the contract expects #010102 — the token system is inverted. Button contrast is 2.32:1.

Hugging Face (56.2%) has no prefers-reduced-motion query. Users who request reduced motion get full animations regardless. It also has --ink and --muted tokens missing — the semantic color system is incomplete.

Mistral (56.5%) has no ::selection rule and no tabular-nums. Its --duration-quick and --duration-slow are both 1000ms — the fast and slow duration tokens are identical, which means the motion timing hierarchy is flat.

The blocked sites

Perplexity and OpenAI both return HTTP 403 to our scoring engine. We've seen this pattern in fintech — Mercury, Coinbase, Revolut, Chime, and SoFi all block automated verification. Now two of the biggest AI companies do it too.

That's 2 of 12 (17%) blocked in AI/ML, compared to 5 of 13 (38%) blocked in fintech. The AI sector is more transparent than fintech — but the two biggest consumer-facing AI companies (OpenAI, Perplexity) are walled off.

The cross-batch pattern

After 4 batches (30 sites + 16 dev tools + 11 fintech + 12 AI/ML = 69 scored sites), the pattern holds:

Zero A-grades across 69 sites. Only Designesy (the reference implementation) scores A.
Infrastructure beats models. Vercel, Modal, and Replicate outscore Cohere, Mistral, and Hugging Face.
Token misconfiguration is the most common failure. --ink set to medium-gray instead of dark. --paper inverted. --signal missing. These are root-level mistakes that cascade through every component.
Button contrast fails on 8 of 12 sites. The most basic accessibility check — can you read the button text? — and two-thirds of AI/ML platforms fail it.
No site uses text-underline-position: from-font or text-decoration-skip-ink: auto. These are resolved tensions from the Cadence typography spec. Zero adoption.