# Verification Release inventory: v1.7.0 · 20 skills · 20 canonical commands · 1 alias · 21 command files · 8 gallery systems. ## BLUF `npm test` verifies the repository, installer, contracts, links, demo surfaces, and the Taste Oracle's internal capture/analysis contracts. Browser QA verifies rendered behavior on the pages and viewports actually tested. Effectiveness is unproven; the terminal blind evaluation has not cleared its release threshold. Release evidence: PASS — receipts current. Effectiveness: BLOCKED. Public copy can be checked independently with: ```bash node tools/release/check-effectiveness-claims.mjs ``` Use both verification layers for a release: 1. **Repeatable local gate:** `npm test` 2. **Manual/browser QA:** render the static pages in a real browser and do the checks below Do not claim "accessible", "zero contrast failures", "all agents auto-trigger", or "all samples are identical" unless the relevant gate below was actually run and the evidence is attached to the change. ## Repeatable Gate Run from the repo root: ```bash npm test ``` `npm test` first runs the Taste Oracle's contract, capture, statistics, judge-quorum, and file-backed CLI tests. It then runs `tools/verify.mjs`, `tools/lint-skills.mjs`, `tools/verify-landing.mjs`, `tools/verify-integration.mjs`, and `tools/verify-gate-audit.mjs`. `tools/verify.mjs` checks: - 20 skill directories exist and each `SKILL.md` frontmatter `name` matches its directory. - Referenced `references/*` and `assets/*` files exist. - Slash commands target real `~/.agents/skills//SKILL.md` paths. - `install.sh --yes` creates canonical `~/.agents/skills` links and detected Claude links. - `install.sh` fails loudly on pre-existing real skill directories instead of nesting stale installs. - Local Markdown/HTML links resolve. - CSS does not contain unspaced `clamp()` math that browsers silently drop. - Demo inputs have accessible names. - Data-viz starter figures expose at least as many table rows as plotted SVG points. - Stale internal skill names such as `dark-mode` are not used as handoff targets. - `a11y-pass/assets/audit.js` parses as JavaScript, avoids `arguments.callee`, and uses browser color rasterization instead of numeric-regex RGB guessing. `tools/lint-skills.mjs` checks the skill *content* (what static link checks miss): dead skill cross-references, design tokens outside the canonical glossary, and frontmatter drift across all 20 skills. `tools/verify-gate-audit.mjs` executes `tastecheck-pass/assets/gate-audit.js` against a minimal fake DOM and asserts its automation contract — `window.__gateAudit` is a `{verdict, fails, warns, notes}` object, the console path still emits, and its output matches the committed golden (`tools/smoke/fixtures/gate-audit-golden.txt`). Scope is deliberate: it guards the attribute-pure checks (a layout-faithful fake DOM would be a partial browser); the layout-dependent stat-band/date-context check is protected by the real-browser fixture in `tools/smoke/fixtures/` instead. `tools/verify-landing.mjs` checks [`index.html`](../index.html), the page connected to GitHub Pages: - Every installed skill has a live `data-skill` surface on the landing page itself. - The page exposes stable operator controls for light/dark/high-contrast themes, the primary CTA path, component states, form validation, empty/error/retry recovery, and keyboard focus. - Form inputs are named, landmarks/headings/live regions/focus/reduced-motion/forced-colors are present. - The data-viz table has one row per plotted SVG point and the chart has title/description text. - The proof section makes the landing page the canonical integration surface instead of sending the primary journey to a side demo. - Landing colors are tokenized with OKLCH and page-level theme wiring is present. `tools/verify-integration.mjs` checks the secondary harness, [`demos/skill-integration.html`](../demos/skill-integration.html): - Every installed skill has a live `data-skill` surface on one local website. - The page exposes stable operator controls for themes, component states, forms, empty/error/retry, and task-list recovery. - Form inputs are named, landmarks/headings/live regions/focus/reduced-motion/forced-colors are present. - The data-viz table has one row per plotted SVG point and the chart has title/description text. - Generic live-copy tells are absent from the improved site. - Responsive/theming foundations are present, including OKLCH tokens and mobile layout guards. ## Browser QA Checklist Run this before publishing visual claims or screenshots: - Open `index.html`, `samples/index.html`, all eight committed `samples/*/index.html`, and every `demos/*.html`. - Check 390px, 768px, and 1280px widths. - Confirm first meaningful content renders, fonts/images resolve, and there is no horizontal overflow. - Open DevTools console; record any errors or warnings relevant to the page. - Paste `skills/a11y-pass/assets/audit.js` and record fails/warnings. - On a FRESH load (no clicks/scroll first), paste `skills/tastecheck-pass/assets/gate-audit.js` and record the verdict + any fails/warns (cold-load state honesty + slop tells). - Do a keyboard pass: skip link, nav, buttons, forms, toggles, and links must be reachable and visibly focused. - Check reduced motion with `prefers-reduced-motion: reduce`. - Spot-check dark and high-contrast theme mappings where a page exposes them. The 1.0 browser receipt automates this matrix for every committed sample and demo at 390px, 768px, and 1280px. It records HTTP/render/overflow/console results, keyboard focus, reduced-motion state, the runnable accessibility and gate audits, and content-addressed screenshots. `verify:release` validates the complete check set, current source digest, and every referenced artifact byte/hash; a pinned `pass` field alone is insufficient. ## Evidence Format For releases, append a short dated note with: - Commit hash. - `npm test` output. - Browser/version used for manual QA. - Pages and viewport widths checked. - Any remaining warnings and why they are acceptable. ## Gate-Auditor Dogfood Date: 2026-06-12 The full `tastecheck-pass` gate was run against the landing page (`index.html`), and it matters *which layer* caught what — the lesson is that the mechanical auditor is the shallow layer. **The mechanical auditor (`gate-audit.js`) passed and was not enough.** On a fresh load it reported 0 cold-load failures (nothing `hidden` defeated by CSS, no error text before input, no content stuck at opacity 0, no stuck skeletons) and one accepted warn — the intentional indigo→violet "before" demonstration (the `deslop-ui` cell's `✕ pill · purple` example, and the `aria-hidden` slop swatch). The reveal-on-scroll already followed the `micro-motion` no-JS rule (`.js .rv` behind a script-added `html.js` hook, with an observer fallback and a `setTimeout` safety net); the display face resolves to "Red0" (Redaction 0, the committed distinctive font) — a NOTE, not a tell. **The structural self-check (`deslop-ui`) caught what the auditor structurally cannot.** The page predated the structural-plane self-check, and applying it by judgment surfaced a real defect the 10-check auditor has no concept of: the section stack was the SaaS **funnel order** — hero → "the problem" → "how it works" → feature grid → proof → CTA. Every section mapped to a skeleton slot, in skeleton order; only the *treatment* was varied. That is exactly the tell `references/structural-tells.md` names. **The fix (this branch):** the landing page was restructured to break the funnel on order *and* count — lead with the working bento (demonstrate before pitching), demote the three-planes section from a problem-opener to a principle that lands *after* the demonstration, and cut the redundant "how it works" section (the interview already lives in bento cell 00). New stack: `hero → skills (live) → what-it-refuses → proof → install`. `npm test` stays green; the auditor re-run is still 0 fail / 1 (accepted) warn; the structural self-check now passes. The takeaway recorded here: a clean mechanical-auditor run is necessary, not sufficient — the structural plane is judgment work, and it bit the pack's own page. ## Historical receipt — 2026-06-06 Date: 2026-06-06 - `npm test`: PASS (`tastecheck verification passed`; `landing verification passed`; `integration verification passed`). - Browser: CloakBrowser Chromium against local `python3 -m http.server` on `http://127.0.0.1:4176/`. - Pages checked in real same-origin iframes at 390px and 1280px: homepage, samples gallery, the eight sample systems, six demos, example-build after page, and `skills/data-viz/assets/chart-starter.html`. - Browser smoke result: 28 page/viewport checks, 0 blank pages, 0 H1-count failures, 0 horizontal-overflow failures, 0 broken images, 0 `a11yAudit()` failures. - Keyboard spot check: first Tab focus lands on the home link with a visible 2px outline. - Homepage `skills/a11y-pass/assets/audit.js` result after the landing-page pass: 0 measured failures and 0 warnings at desktop 1440px and mobile 390px. Historical landing-page 20-skill evidence (receipt dated 2026-06-06): - Page: `index.html`, served as the GitHub Pages homepage. - Static gate: all 20 installed skills are represented by `data-skill` on the homepage. - Real operator paths covered by the page contract: light/dark/high-contrast themes, primary CTA, component loading/disabled/default states, invalid and valid email validation, empty/error/retry recovery, chart/table parity, keyboard-focus target, and live-region status updates. - Proof copy now identifies the homepage itself as the integration surface. - Visual checks: desktop hero/nav and mobile hero/skills/proof sections inspected by screenshot; the collapsed desktop nav, mobile tiny utility text, palette overlay collision, and mobile install command overflow were fixed during the pass. Secondary integration harness evidence: - Page: `demos/skill-integration.html`. - Desktop 1440px and mobile 390px: title/URL correct, 20 unique skills exposed, 0 horizontal overflow. - Real operator paths executed in browser: light/dark/high-contrast themes, component loading/success/disabled, invalid and valid form submission, empty/error/retry task-list states, and chart/table parity. - `skills/a11y-pass/assets/audit.js`: 0 measured failures, 0 warnings on desktop and mobile. - Manual keyboard check: skip link, nav link, and theme button receive visible 3px focus outlines. ## Current evidence — 2026-09-14 - Repo at the 4f3c48c-era surface-truth pass (this PR): the mechanical battery is green (33 eval tests + gate-audit + browser lane; receipts: hq `docs/cto/2026-09-14-tastecheck/SKILL-BATTERY-VERIFY.md`). - Release receipts remain stale-until-rerun pending `npm run finalize` (which this PR runs); the projected release status above stays UNVERIFIED and Effectiveness stays BLOCKED until the registered producers rerun at the final source.