@@ -41,15 +41,14 @@ jobs:
         run: node visualize/scripts/check-library-policy.mjs --quiet
 
       - name: Detector smoke — every template passes --strict
-        # Runs detect.mjs against every template.html (canonical + variants)
-        # in --strict --fast mode. Any error-severity finding fails the build.
-        # The canonical templates ARE the reference; if a template fails the
-        # detector, the rule set or the template is wrong.
+        # Runs detect.mjs against every template.html (canonical + variants).
+        # Any error-severity finding fails the build. Includes WCAG contrast
+        # math via the vendored culori library.
         run: |
           for f in $(find visualize/templates -name '*.html'); do
-            if ! node visualize/scripts/detect.mjs --strict --fast "$f" > /dev/null; then
+            if ! node visualize/scripts/detect.mjs --strict "$f" > /dev/null; then
               echo "detector failed on $f"
-              node visualize/scripts/detect.mjs --strict --fast "$f"
+              node visualize/scripts/detect.mjs --strict "$f"
               exit 1
             fi
           done
 
@@ -22,24 +22,44 @@ Drop a new directory under `visualize/themes/<slug>/` with `tokens.css` (CSS cus
 
 ## Anti-slop detector
 
-Templates and (where applicable) brand-overlay-applied artifacts must pass the deterministic anti-slop detector:
+Templates must pass the deterministic detector:
 
 ```sh
-node visualize/scripts/detect.mjs --strict --fast path/to/artifact.html
+node visualize/scripts/detect.mjs --strict path/to/artifact.html
 ```
 
-26 rules across 5 categories: `fossil` (lorem-ipsum, mock identity, citation artifacts, sycophant footers, stock LLM openers), `slop` (system-default fonts, generic gradients, gradient text, AI palette, dark glow, icon-tile stacks, emoji headings, center-everything, monotonous spacing, bounce easing, uppercase body, side-tab), `a11y` (missing alt, missing lang, empty link, low contrast), `meta` (missing title, favicon, OG tags, external script not on allowlist), `perf` (layout-thrash transitions).
+~23 rules across 5 categories — all mechanical patterns regex or `node-html-parser` can identify reliably, plus WCAG contrast computed via the vendored `culori` library:
 
-CI runs the detector against every template file in `--strict --fast` mode; any `error`-severity finding fails the build. `--fast` skips the contrast check (which needs `jsdom`; ships as info-skip otherwise).
+- **fossil** — lorem-ipsum, citation artifacts (turn0search0 / `<!-- claude: ` / 【N†source】 / ASSISTANT:), AI-attribution disclaimers ("Generated by Claude / with AI assistance")
+- **slop** — generic gradients (purple→pink / indigo→cyan), gradient text, dark glow (text-shadow with high blur + chroma), glassmorphism (backdrop-filter + translucent), emoji headings, monotonous spacing, bounce easing, side-tab, system-default fonts (when brand-aware)
+- **a11y** — missing alt (including lazy alts like `alt="image"` / `alt="DSC_0042"`), missing/empty lang, empty links, WCAG-AA contrast, heading structure (missing h1, multiple h1, skipped levels)
+- **meta** — missing title, missing favicon, missing OG / Twitter card, external script outside the library-policy allowlist
+- **perf** — layout-thrash (transition on width/height/top/left, plus `transition: all`), img without width/height (CLS risk)
 
-Output formats: human-readable default, `--json` for NDJSON. Programmatic use:
+CI runs `node detect.mjs --strict <file>` against every template; any `error`-severity finding fails the build.
+
+**Rules deliberately NOT in detect.mjs** — vocab-unbounded fossils (mock identity beyond the obvious / slop openers / sycophant footers / imperative-tricolon), brand-and-register-dependent slop (ai-color-palette / icon-tile-stack / center-everything / uppercase-body), and structural-pattern slop the detector can't distinguish from legitimate design (triple-feature-card / trust-signal slop / scroll-jacking / responsive overflow). Those live in the `/visualize:review` LLM-judge prompt in `SKILL.md` § "Review verb procedure" — the agent reads them when reviewing an artifact and applies judgment the detector can't.
+
+The split is deliberate: mechanical patterns get detector rules (cheap, deterministic, CI-enforced); judgment-shaped patterns get prompt instructions (require register awareness, brand context, semantic understanding).
+
+### CLI
+
+- Output: human-readable default; `--json` for NDJSON.
+- `--strict` — exit 2 on any error finding.
+- `--skip slop/emoji-heading,a11y/missing-alt` — comma-separated rule IDs to skip.
+- `--config .visualize-detect.json` — config file (severity overrides + skip list).
+- `--brand DESIGN.md` — explicit brand profile path (auto-loaded from any ancestor of the artifact otherwise).
+
+### Programmatic
 
 ```js
 import { detect } from './visualize/scripts/detect.mjs';
-const findings = await detect({ path: './artifact.html', skip: ['slop/emoji-heading'] });
+const findings = await detect({ path: './artifact.html', skip: ['slop/emoji-heading'], brandProfile: { fonts: ['sohne'] } });
 ```
 
-Configuration via `.visualize-detect.json` at the project root (severity overrides, skip lists). See the spec § "Anti-slop detector backs the `review` verb" for the rule contracts and the F103 audit history.
+### Vendored dependencies
+
+`detect.mjs` imports two libraries from `visualize/scripts/vendor/`: `node-html-parser` (real DOM walking) and `culori` (WCAG contrast). Both bundled inline (no `npm install` required after `npx skills add display-dev/visualize`). See `visualize/scripts/vendor/README.md` for re-bundle instructions when upstream releases.
 
 ## Library / CDN policy
 
 
@@ -105,29 +105,54 @@ The combination of universal HTTP fallback + unauthenticated publish path means
 
 ## Review verb procedure
 
-`/visualize:review <path>` always runs in two layers:
+`/visualize:review <path>` runs in two distinct layers — deterministic checks via `detect.mjs`, then LLM judgment via the rules below. Each layer catches what the other can't reliably see.
 
-**Layer 1: deterministic findings via `detect.mjs`.** The agent shells out to:
+### Layer 1 — deterministic findings via `detect.mjs`
 
 ```sh
-node $SKILL_DIR/scripts/detect.mjs --json --fast <path-to-artifact.html>
+node $SKILL_DIR/scripts/detect.mjs --json <path-to-artifact.html>
 ```
 
-NDJSON output, one finding per line plus a trailing summary line. Each finding has `{ ruleId, category, severity, locator, message, snippet?, suggestion? }`. Severities: `error` / `warn` / `info`. Categories: `fossil` / `a11y` / `meta` / `slop` / `perf`. The 26 rules cover the failure modes the agent reliably reaches for — gradient text, AI palette, generic gradients, dark glow, fossil openers, citation artifacts, sycophant footers, accessibility basics, publishing-hygiene meta tags. Errors are ship-blockers per universal law 3; warns + infos are taste-calls the agent triages.
+NDJSON: one finding per line + trailing summary. Each finding carries `{ ruleId, category, severity, locator, message, snippet?, suggestion? }`. The detector auto-loads `DESIGN.md` from any ancestor of the artifact path so the brand-aware rules (currently `slop/system-default-font`) work; pass `--brand <path>` to override.
 
-`--fast` skips the contrast check (which needs jsdom; ships as info-skip otherwise). Drop `--fast` for the full check when jsdom is installed.
+The detector ships ~23 rules — all of them mechanical: HTML/CSS patterns regex can identify or that `node-html-parser` walks reliably, plus WCAG contrast computed via `culori`. Categories: `fossil` (placeholders, citation artifacts, AI-attribution disclaimers), `slop` (gradients, gradient text, glow, glassmorphism, emoji headings, monotonous spacing, bounce easing, side-tab, font when brand-aware), `a11y` (alt, lang, empty link, low contrast, heading structure), `meta` (title, favicon, OG, external script), `perf` (layout thrash, missing img dimensions).
 
-**Layer 2: LLM judgment on top.** Once the agent has the deterministic findings, it adds the qualitative reading the detector can't see:
+### Layer 2 — LLM judgment in this prompt
 
-- Brand-profile-overrides per universal law 1: a `slop/system-default-font` finding on a brand that explicitly declares Inter is a category error — demote to info.
-- Aggregate-as-tell per universal law 4: 6 `slop/center-everything` info findings = ship-blocker; 1 = taste-call.
-- Genre-reflex check: did the artifact fall into its template's stock aesthetic (red banners on Postmortem; purple gradients on AI essay)?
-- Hierarchy + scanability: does the eye land in the right place first?
-- The 6 universal laws as the override grid for any single-rule finding.
+Things detect.mjs deliberately doesn't try to catch — register-dependence, vocab-unboundedness, semantic-structure calls. The agent must read the artifact and assess:
 
-**Output to the user:** prioritised list, errors first, with the brand-overlay context applied. The agent never just dumps the NDJSON — it interprets.
+**Fossil-shape failures** (vocab is unbounded, so no detector list):
 
-`--strict` exits 2 on any error finding (for CI consumers). The `/visualize:review` verb itself never sets `--strict` — it's an audit, not a gate.
+- **Mock identity** beyond the obvious — fictitious customer quotes ("Sarah Johnson, VP Marketing"), placeholder phone / address / company / brand names that aren't in the obvious set (Acme, Globex, Anytown, TBD, XXX-XXX-XXXX).
+- **Slop openers** — "leverage / streamline / empower / unlock / robust / holistic / seamless / cutting-edge / In conclusion / TL;DR:" packed into the opening paragraphs.
+- **Sycophant footers** — "I hope you find this useful", "Happy to clarify", "Hopefully this addresses your needs", "Best, Claude", anything signed by an AI.
+- **Imperative tricolon** ("Ship faster. Build smarter. Scale forever.") — three commands ending in periods, regardless of content. Can't be redeemed.
+
+**Brand- and register-dependent slop** (mechanical detector would false-positive):
+
+- **AI color palette** — is the primary / accent stack a generic purple-pink-cyan trio, OR is it the brand's actual identity?
+- **Icon-tile-stack** — every section heading has a rounded-square icon above it. Genuine AI section-marker tell, or is this a deliberate icon-led design system?
+- **Center-everything** — body prose centered, section after section. False positive for Bauhaus / Swiss-poster / academic preprint registers; real signal for AI-essay register.
+- **Uppercase body** — long all-caps passages. False positive for brutalist / letterpress / show-bill registers; real signal otherwise. Acronyms (HTML / POST / REST / WCAG) don't count.
+
+**Structural-pattern slop** the detector can't reliably distinguish from legitimate design:
+
+- **Triple-feature-card** — 3-column grid of `<h3>adjective + noun</h3>` blocks with one paragraph each. The AI-landing canonical pattern.
+- **Trust-signal slop** — "Backed by leading investors", "10,000+ teams", template testimonial stack.
+- **Infinite animations / parallax / scroll-jacking** — sometimes deliberate, often not.
+- **Responsive overflow** — fixed min-width above 320px; `100vw + padding`; long button text that mobile-clips. Detector can't simulate viewport.
+
+### Override grid (apply to BOTH layers)
+
+- **Universal law 1** — brand profile overrides any single-rule finding. A `slop/system-default-font` hit on a brand that declares Inter is a category error; demote.
+- **Universal law 4** — info-severity rules ship-block in aggregate. 6 `slop/emoji-heading` infos across a presentation = pattern; 1 = taste-call.
+- **Universal law 6** — the reader is the judge. Does each finding make the reader's job harder, or just satisfy a rule?
+
+### Output to the user
+
+Prioritised list. Errors first (detector + the unbounded-vocab fossils the agent catches in layer 2). Warnings + info aggregated by category, with the brand-overlay context applied. Never dump raw NDJSON; always interpret.
+
+`--strict` exits 2 on any error finding — for CI consumers, not the `/visualize:review` verb itself.
 
 ## Brand profile