Built to be usable by everyone. Here is what we have checked.

Automated audit of every page

Every route in both languages is run through axe-core with its full WCAG 2.0, 2.1 and 2.2 rule set at levels A and AA, plus the rule that requires landmarks to have distinct names. The audit checks that it evaluated at least sixty rules on each page, because a narrowed audit reports the same green as a full one and this project once shipped exactly that. A clean run is the floor, not the standard: automated checks find only part of what matters.

Keyboard, focus and structure

In a real browser, the suite checks that the skip link is the first tab stop and moves focus to the main content, in both languages; that every navigation item is reachable by keyboard in order; that a focused control and a link target are not hidden behind the header; that every focus indicator in the header and the footer is painted against the ground it sits on, measured in pixels; and that every page has exactly one first-level heading and skips no heading level.

• Skip link first, and it works.
• Every focus ring measured for contrast.
• One first-level heading per page, no level skipped.
• Zoom never disabled, and the page language declared.

Small screens, large text and reflow

The browser suites run at desktop, wide-desktop, tablet and phone widths — 1440px, 1280px, 820px, 393px and 360px — and a separate project runs at 320px, the width WCAG 1.4.10 Reflow names. At every width, every page is measured at normal text size and at 200% text size, and must never scroll sideways. A further check measures every element on every page and fails if any one renders wider than the box that holds it.

Reduced motion

With the reduced-motion preference set, every element on every route is measured and must have no perceptible animation or transition; the same route is measured again without the preference to prove there was motion to suppress. Motion collapses to its end state rather than being deleted, so nothing that teaches through animation stops teaching.

Colour contrast

The contrast of every text colour on the surface it is named for, and of every focus ring on its own surface, is computed from the design tokens by a test rather than transcribed into a table. A token that fails the minimum fails the build.

The person experience, before it is public

Every route of the person experience renders and can be operated with JavaScript turned off, and the release gate runs it that way. Two things are deliberately not offered without scripting — the question about which topic matters most, and recording anything at all — and the suite asserts both as absences rather than letting them pass as working. Register entry A-22. It is audited at 320px on every one of its routes, in both languages, and the crisis resources it ships are in the markup rather than loaded afterwards.

A border colour with no headroom

One border colour used on cards measures 3.32 to 1 against the editorial background, where the non-text contrast minimum is 3 to 1. It passes, by 0.32, and no test pins that figure, so a change to the surface could erode it unnoticed. Register entry A-10.

A focused navigation item that can be partly clipped

At desktop widths the header’s navigation row is still a horizontally scrolling container, and a keyboard user tabbing into it can leave a focused item partly outside its edge — never entirely hidden, which is the level-AA requirement, but partly, which the enhanced level forbids. Below 1280px — and at every width once text is enlarged — the destinations move into the menu instead and there is nothing to clip. The sweep that found this has not been re-run against the wrapped row, so the entry stays open until it has. Register entry A-20.

Space does not grow when a reader enlarges the text

Every size of text on this site is set in units that follow the reader’s own text size, so raising it to 150% or 200% enlarges the words and nothing is cut off, hidden or pushed sideways — we measure that at every width we build for, including 320px. What does not follow is the space: padding, gaps and the page margin are fixed lengths, so as the words get bigger the room around them stays where it was. At 200% text the space around a card’s copy ends up narrower than the copy it surrounds, where at the ordinary size the design gives it half as much again. Nothing becomes unreadable and nothing is lost; the page gets tighter than it was drawn to be, and a reader who needs large text gets the least generous version of it. Fixing it means changing the unit every spacing in the system is written in, which touches every page HABLA has, so it is its own piece of work with its own review rather than something done in passing. Register entry A-49.

Escape does not close the header menu on narrower screens

Where the header’s menu is a disclosure — below 1280px at any text size, and at every width once text is enlarged — it is a native one and this site ships no script for it, so pressing Escape does not close it. Nobody is stuck: one Shift+Tab from the first link reaches the control that opened the menu, where Enter closes it, and tabbing forward leaves the header altogether. The menu also gives way when keyboard focus reaches the page behind it. What is open is the keystroke a reader may reasonably expect, and it stays open until a screen-reader session tells us whether the control already reads as a way out. Register entry A-36.

A long table region whose focus ring lands off the screen

One page for leadership carries a table region you can scroll, and it is taller than the screen it appears on. It is reachable by keyboard, as it must be, and on a phone the browser scrolls it into view in a way that puts its top edge and its focus ring above the visible area. It is never entirely hidden, so the level-AA requirement holds; the enhanced one does not. Register entry A-30.

The chat panel closes when focus leaves it, and only the keyboard path was measured

HABLA AI’s panel closes when focus moves out of it, so a floating panel never sits over a control you have just focused. That was measured with a keyboard, at every width the suite runs. It has not been measured with a screen reader, where somebody reading past the panel could find it closed behind them. Register entry A-28.

The chat’s disclosure scrolls, and on a small phone the only sign is the cut at its edge

On a 360px-wide phone the chat’s privacy disclosure fills the panel and the cut at its lower edge is the only sign there is more to read. A reader who scrolls reaches all of it, and the box to type in is below. Register entry A-29.

The chat launcher still sits over page content on wider screens

Below 1280px the chat launcher sits in the flow of the page and covers nothing. Above that it floats in the corner, and at the top of the page it still overlaps text on up to 11 of this site’s 46 pages, measured at each width we build for. On every one of those it is the tail of a paragraph or one line of a list, and never more than a fifth of it. What is open is that a floating control and a page that fills its own margins cannot both have that corner. Register entry A-34.

Two rules that read this website only

The scan for styles nothing uses and the byte budget for each page still read only this website and not the person experience. The person experience was measured by hand and was clean; what is open is that nothing would catch it regressing. Register entry F-28.

A clearance nothing measures at very large text

The room the sticky header leaves above a focused control is asserted by test at normal and at double text size, and at one width only when text is enlarged further. It was measured by hand at those larger sizes across every width we build for and was correct; what is open is that no test would catch it changing. Register entry F-36.

The sticky header grows taller when text is enlarged

At 1280px and 1440px with text at 200%, the header folds to its Menu control — the same bar a phone gets — so the sticky bar stays 65px instead of growing to 95px, and every destination stays reachable. Until 14 September this clause said the row wrapped instead, and that nothing on it overlapped. The height was right; the second half was not. It had been measured between the header’s blocks rather than between the links inside them, and measured link by link there were 202px of navigation in English and 272px in Spanish with nowhere to go, with the Contact control painted underneath the language switch. That is fixed, and it is written here because the wrong sentence was published. What is still open is the reason this clause exists: the test pins the bar at exactly 65px and runs at the normal text size only, so nothing in the repository would notice it growing. This is the desktop half of the gap in the clause above. Register entry A-38.

A focus failure reported in review that our own tests could not reproduce

Reviewers reported that at 360px, 393px and 820px wide, with text at 200%, a focused link in the header was completely hidden. Walking the same pages with the same key presses under this site’s own test devices, every link measured fully visible — on the build before the fix as well as the one after it. The narrower case they also reported, at 320px, did reproduce and is fixed. What is open is that we cannot yet say how the other three numbers were produced, and writing that down is better than dropping it. Register entry A-37.

How a screen reader says P²HI^™ has not been tested

The short name for the category this site introduces is written P²HI, with a superscript. Screen readers usually read that superscript aloud. Some voices, at some settings, skip it and say “P H I” — which in healthcare and government means Protected Health Information, a different thing entirely, and a spelling this site does not publish. We have not yet listened to it on a real screen reader, so we are not claiming it is fine. Wherever the short name appears the full phrase is next to it, so a reader hears Privacy-Preserving Human Intelligence™ in the same breath. Register entry A-73.

With no script, the way back to the form opens an empty one

In the person experience, a message that cannot be sent leads to a screen offering a way back to the form. Where a script runs, it returns the person to what they wrote. Where no script runs it is an ordinary link, so it opens a fresh form with every box empty. Nothing the person wrote is destroyed — it is one more step back in the browser — but the screen does not say so, and WCAG 2.2 asks at level A that somebody not be made to enter the same thing twice. The person experience is not public yet. Register entry A-33.