Scope declared up front: this is the deterministic guardrail layer (computed WCAG contrast). The perceptual “does it look premium” verdict is unsolved even at frontier SOTA (best VLM ~72.5% on design-aesthetic judgment; the leading UI-quality benchmark uses human expert votes, not a model) and is filed for a partnership VLM/human judge — it is not faked here.
The state-of-the-art bar (primary-sourced)
Five axes define “experiential quality.” Each bar below is cited to a primary source so we measure against the real external ruler, never one we invented.
| Axis | SOTA bar | Source |
|---|---|---|
| Performance (Core Web Vitals) | LCP ≤ 2.5s · INP ≤ 200ms · CLS ≤ 0.1, at p75 (INP replaced FID in 2024) | web.dev/vitals |
| Accessibility contrast | Normal text 4.5:1 · large (≥24px / ≥18.7px bold) 3:1 · AAA 7:1; exact sRGB-luminance formula | W3C WCAG 2.2 §1.4.3 |
| Accessibility conformance | WCAG 2.2 AA (86 SC); automation covers ~57% of issues / ~30–40% of criteria — a floor, not proof | W3C WAI · Deque |
| Design system / emitting | W3C design tokens ($type/$value); 3-tier tokens; OKLCH color; fluid type; container queries; states + reduced-motion | designtokens.org · web.dev |
| Visual & structural testing | Deterministic pixel-diff (YIQ+AA-ignore), ARIA-tree snapshots, Zhang-Shasha tree edit distance; console/CSP capture | Playwright · Chromatic |
| Perceptual “premium” | UNSOLVED — rubric-anchored VLM as a weak proxy calibrated to human pairwise votes (TrueSkill) | AesEval-Bench · UI-Bench |
The gap map (computed, full-scale)
Every live surface fetched over sovereign TLS and audited by nx_uiq_contrast — real WCAG 2.x contrast on the actual colors, scope-aware for light/dark themes. Findings a presence-checker cannot see:
| Surface | Verdict | Computed finding |
|---|---|---|
| / & /sitemap | FAIL | dark-mode accent --ac on --bg = 2.34:1, on --soft = 2.18:1 (accent has no dark variant) |
| /finance | FAIL | --nx-color-ink2 on --nx-color-surface2 = 4.39:1 — sub-threshold (<4.50); only computation catches 4.39 vs 4.50 |
| /compare · /research · /roi · /uidebt | PASS | tokenized, all text×surface pairs ≥ 4.5:1 in both themes |
| /sheriff | no tokens | hardcoded background:black; color:gainsboro — zero design tokens, dev-grade |
| /search · /video · +12 more | no tokens | raw #hex colors (44–83 per page), design system not adopted |
Honest envelope: v1 computes contrast for declared --custom-property color tokens (#hex/rgb()/named), paired by role (text×surface), per resolved theme. In-rule color:#x pairing, non-text (3:1) & large-text tiers, oklch()/hsl() parsing, and rendered element-on-background stacking are filed rungs.
The new sovereign capability
nx_uiq_contrast is a live NAS-registered MCP tool: it parses any page, extracts every color token, and computes the exact WCAG 2.x contrast ratio — the W3C’s formula, KAT-proven against its reference values (black/white = 21.00, #767676/white = 4.54). It replaces “does the CSS feature appear” (presence) with “does the page conform” (measurement). It also seeds a shared color/WCAG library (nx_uiq_color) that collapses three duplicate luminance implementations into one.
Why this matters
Our pages scored well on a ruler that checked whether a token existed once. This one checks the actual numbers — and immediately found a dark-mode accent failure and a 4.39:1 near-miss no human would catch. That is the difference between hygiene and conformance.
The road to SOTA (filed rungs & ownership)
| Rung | What | Owner (RACI-A) |
|---|---|---|
| U-fix | Add dark-theme accent variant ≥4.5:1; fix /finance ink2 near-miss | racing / emitter |
| U3 | Retrofit the ~15 dev-grade surfaces onto nishi_ds + nx_ui_kit | racing / owners-per-surface |
| UIQ v1.1 | In-rule color:#x pairing + non-text (3:1) + large-text tiers + oklch()/hsl() | experiential |
| UIQ-perf | Static Core-Web-Vitals predictor (transfer weight, render-blocking, unsized-media/CLS, connection reuse) | engineer-browser |
| UIQ-struct | ARIA-tree + Zhang-Shasha regression, console/CSP capture during crawl | experiential |
| UIQ-perceptual | Rubric-anchored VLM judge calibrated to human votes (partnership; will not fake) | modelwright / referee |
| Consolidate | Merge nx_contrast + nx_ui_contrast + nx_brand_guard math onto nx_uiq_color (OO, via nx_oo_extract) | architect |