Set the web standards as data and check every build against them

Context and Problem Statement

Of the five real sites, three declare hreflang (one always pointing at the home pages), four a canonical, two a sitemap, two structured data, one a skip link; one has no focus styles, one loads fonts from Google, two load a reCAPTCHA script; three keep partial checkers of their own (i18n parity, contrast, dead links); none has continuous integration, an accessibility tool or Intl formatting. All five keep Spanish at the root and only one translates slugs, and even it keeps Spanish for its legal and thank-you pages in every language. What quality bar does every site meet, and how is it checked rather than hoped for?

Decision Drivers

Criteria 1 (trust), 4 (accessibility, WCAG 2.2 AA), 7 (three languages) and 9 (generated and tested) of ADR-0009.

Considered Options

• The standards as data (brand/standards.yaml) and a checker that runs them, with the anatomy’s patterns, over any built site; this site the first to pass
• A browser-based audit (axe, Lighthouse) run by hand before each deploy
• A written standard with no automated check

Decision Outcome

Proposed: the first option. brand/src/check-site.ts reads standards.yaml and anatomy.yaml, walks a built site’s HTML and lists every violation; pnpm check runs it over this site’s dist/ and fails the build on one. Running it on this site found 984 violations and caused the fixes below; it now passes 26 rules over 175 pages with none.

• Accessibility. WCAG 2.2 AA is the floor. Checked statically: lang on html equal to the page’s locale; exactly one h1 and no skipped heading level; exactly one main and a skip link to it; alt on every image; a name (role="img" and a title or label) or aria-hidden on every inline SVG; text or a label on every link, no bare #; a label and autocomplete on every form control. Focus (--focus-ring), target size (24 px, controls 44 px) and reduced motion are the shape tokens’ job (ADR-0021); contrast is proven by the colour build (ADR-0015). Keyboard operation, zoom to 400 % and error announcement stay manual and are checked on each shared component when it is built.
• Addresses and languages. Every page under its language (/es/, /en/, /pt/), every segment translated, lowercase ASCII with hyphens and a trailing slash; the root redirects by Accept-Language with Spanish as the fallback (this site’s Worker already does), x-default points at the root; old addresses redirect permanently. hreflang for the three languages plus x-default, absolute, the page itself included; a canonical that is the page’s own address. Numbers, dates and units through Intl in the page’s language (es-AR, pt-BR, en).
• SEO on indexable sites: title 10 to 70 characters, description 30 to 160, og:title, og:description, og:url, og:image (the brand’s profile picture at 1200 px unless the page has its own), Organization JSON-LD on the home page naming AKRIBIS Group as parentOrganization, a sitemap with every page. Brand is noindex while it is a draft and carries all of it but the sitemap; its home page’s JSON-LD lists the nine units and AKRIBIS One from the register.
• Budgets per page, uncompressed: HTML 200 kB, CSS 80 kB, JS 150 kB, fonts 260 kB, at most three preloaded fonts; width and height on every image; fonts from Brand only, never from a font CDN. On this site the budget turned the inlined study sheets and marks into files: the Colour page fell from 656 kB to under 200 kB, the 429-row download list became the SVG masters with a pointer to manifest.json.
• Consent. No third-party script before consent; a script from an analytics or advertising host must be gated (data-consent). The default is none: where nothing tracks, no cookie notice. Cloudflare’s cookieless beacon is the one analytics allowed without it.
• The checker. node brand/src/check-site.ts <dist> --kind <unit|parent|one> prints every violation grouped by rule, the budgets of the largest page, and exits with an error; it reads standards.yaml or a site’s own copy. /standards.json and /anatomy.json publish the rules for any tool. Every site runs it before deploying; this repository runs it in pnpm check.

Consequences

• Good, because the standard is the test: a page that skips a heading level or loads a font from a CDN fails the build instead of a review.
• Good, because the same rules, as data, serve twelve sites and an agent reading llms.txt.
• What a browser sees (computed contrast, focus order, keyboard traps) is not checked here; browser-based checks come with continuous integration, which no site has yet.
• Open: image formats and the LCP’s size; the redirect tables that move the sites’ current addresses to the grammar.