designesy.org/benchmarks3 toolsresearched 2026-08-01self-score 100.0 Athe methodology

Benchmarks

Three tools, three questions. hallmark stops slop as a design is generated, slop-eval scores slop after the fact, and designesy verifies a site against a versioned contract. Where they overlap, and what each one catches alone.

View as Markdown

The three tools

designesy from its registry; the others as researched on 2026-08-01

designesy, hallmark and slop-eval compared, attribute by attribute.
Attribute	designesy	hallmark	slop-eval
Question it answers	Does this conform to the contracted design system?	Does this look AI-generated?	How much slop does this design contain?
Kind	Contract conformance verification	Anti-slop gate at generation time	Anti-slop evaluation skill
Delivery	URL API and MCP tools an agent can call	Agent skill (a prompt-encoded rule set)	Agent skill (a deterministic scoring script)
Checks	45 in 14 categories	57 binary gates and 6 pre-emit axes	108 tells in 6 families and 2 positive axes
License	See designesy.org	MIT	Apache-2.0
GitHub stars, 2026-08-01	n/a	20.5k	40 (parent repository)

A full pipeline can use all three: hallmark while a design is generated, slop-eval to score what exists, designesy to verify the result against the contract. Of the three, only designesy scores a URL through an API, which is what CI, leaderboards and automated pipelines call.

Where they overlap

8 designesy checks with an equivalent in either tool

designesy checks with an equivalent in hallmark or slop-eval.
Check	What it asks	hallmark	slop-eval
`v06, v22`	Contrast (WCAG AA and APCA)	gates 40, 41	X5 (critical)
`v03`	Focus-visible rings	gate 26	X14
`v05`	prefers-reduced-motion	gate 27	none
`v11`	No transition: all	gate 10	none
`v28`	Reading width 45 to 75ch	gate 25	none
`v26`	At most 3 font families	gate 37	none
`v02`	No horizontal scroll	gate 34	none
`v24`	Touch targets 44px and up (WCAG 2.5.5)	gate 26, in part	none

What designesy checks that they do not

20 checks with no hallmark or slop-eval equivalent

v01Token values match the live :root foundation (contract to live conformance)
v08, v09, v10Poise and Takt verification: the interaction-feel contract
v12will-change kept to transform and opacity
v13Press scales 0.96 and 0.985, above the 0.95 floor
v14Cadence typography rules match the contract
v15Font smoothing: antialiased and grayscale
v16A rem-based scale with a 16px root
v17Line-height by role (headings 1.08, body 1.55)
v18text-wrap: balance and pretty both present
v19tabular-nums in 8 places or more
v20::selection styled in the brand accent
v21Core Web Vitals (LCP 2.5s or less, INP 200ms or less, CLS 0.1 or less), measured by the full audit in a browser; the static score marks it manual
v23Duration tokens, --duration-quick to --duration-slow
x01, x02, x03font-synthesis, text-underline-position and skip-ink
v27Input font size of 16px or more (no iOS zoom)
v29Token layers, primitive to semantic to component (DSAF A1.1)
v34AI-disclosure readiness (EU AI Act Article 50, which applies from 2 August 2026)
v35Forced-colors readiness
v36Unicode security: UTS #39 confusables in token names and CSS
v37DESIGN.md spec layer (DESIGN.md format lint)

What they check that designesy does not

the gaps: how a design looks, which the contract leaves to them

AI-slop look: purple gradients, centred heroes, fake chromehallmark 1 to 7, 42, 43, 45, 47; slop-eval C1 to C15, T1 to T10, K1 to K27, L1 to L21
Structural variety across pageshallmark 8, 20, 21, 32, 57; slop-eval L15, L19, L21, axis 7
Signature and uniqueness, a positive rubricslop-eval axis 7 (a 7-element formula, weighted 3 times)
Cohesion, a positive rubricslop-eval axis 8 (4 checks)
Invented metrics and fake copyhallmark 46; slop-eval W3
AI fingerprints in navigation and footerhallmark 42, 43; slop-eval L14, L20, K12
Re-drawn UI chrome: fake browser bars, phone frameshallmark 47; slop-eval K7, K16
Token improvisation: inline hex that bypasses the token blockhallmark 48; slop-eval C9, C10
Hover boop, hover lift and hover scalehallmark 11; slop-eval M2, M4
Dead controls and fake interactivityslop-eval M8 (critical)
Pre-emit self-critique on 6 axes scored 1 to 5hallmark

On designesy.org itself

weekly run of 2026-10-12

Each tool on designesy.org.
Tool	Result	What it found
designesy	A100.0	38 pass, 3 warn, 0 fail, 0 skipped. 3 manual checks (v02, v04, v21) need a browser, so they sit outside the score.
hallmark	Not run	A skill with no URL API. Expected to flag near-pure black and white, zero-chroma neutrals and the navigation structure.
slop-eval	Not run	A skill with no URL API. Expected to flag the cool charcoal dark theme, the mono house voice (minor), and to pass the focus states.

The wider field

newer entrants worth tracking, as of 2026-08-01

Newer design-verification tools, as of 2026-08-01.
Tool	Approach	What sets it apart
Impeccable	46 AI-slop tells, in the slop-eval lineage	The largest tell catalogue, community-sourced
design-slop-cop	14 anti-slop patterns, rule-based	Light and fast, with a focused set
anti-slop-design	Design-quality guardrails	Lints before output instead of scoring after
Atlassian ADS MCP	A design-system MCP server (v0.21.1)	Benchmarked against design.md: fewer tokens, lower variance

mcpservers.org indexed 81 design-related MCP servers on 2026-08-01. The designesy MCP server has 17 tools; see the MCP docs.

Sources

designesy score: the weekly leaderboard run of 2026-10-12, engine 1.3.0
hallmark, 57 gates: github.com/Nutlope/hallmark, skills/hallmark/references/slop-test.md
slop-eval, 108 tells: github.com/fabricioctelles/skills, skills/slop-eval/references/tells.md
slop-eval scoring: github.com/fabricioctelles/skills, skills/slop-eval/SKILL.md
slop-eval upstream: pols.dev/slop.md
@google/design.md: npmjs.com/package/@google/design.md (v0.4.0, Apache 2.0)