Accessibility verification

A clean automated scan proves that a set of machine-checkable rules found nothing. It says nothing about whether a name is meaningful, an order is logical, an announcement is useful, or a flow can be completed without a mouse. Those need a person driving the interface.

When this fires

Before any claim that a screen, component or flow is accessible, keyboard-operable, screen-reader usable or WCAG-conformant — your own work or someone else's, and when re-checking after fixes. It does not fire to write the fixes; that is accessibility.

Procedure

1. Fix the scope in writing first. Which routes, which states (empty, loading, error, logged in), which viewport, which browser. A claim without a named scope is not a claim — everything outside it stays explicitly unverified.
2. Render the real thing. The running app at the route that changed, not the source, not Storybook alone if the route composes differently. A stale build verifies nothing.
3. Run an automated scan and record it exactly. Use whatever the project already has — an axe-based checker, the browser's accessibility audit, a CI accessibility step. Record the tool and version, the URL, the rule ids that failed and the count. Treat the result as a floor: scanners find only a minority of real barriers, and they cannot judge meaning or order at all.
4. Do the keyboard pass with the pointer unused. Tab from the top: write down the order, and whether it matches the visual order. Check the skip link, every control being reachable and operable with Enter/Space/arrows, Escape closing overlays, no trap, focus returning after a dialog closes, and the focus indicator being visible and not covered by sticky headers or banners.
5. Do the screen-reader pass on at least one real pairing — VoiceOver with Safari, or NVDA with Firefox. Read the flow start to finish, then complete the task with the screen buffer as your only information. Check: does every control announce a name, a role and its state; do headings and landmarks describe the page; do errors and status changes get announced when they happen. See references/screen-reader-passes.md for how to drive each one.
6. Check zoom, reflow and text spacing. 200% browser zoom, and a 320 CSS px wide viewport: no horizontal scrolling, no content lost, no overlap. Then apply the text-spacing condition (line height 1.5×, paragraph spacing 2×, letter spacing 0.12em, word spacing 0.16em) by injecting CSS, and confirm nothing is clipped or overlapped.
7. Measure contrast on the rendered pixels, in every state — default, hover, focus, selected, error, and text over images or gradients. A token that passes in the palette can fail in place.
8. Reach the non-happy states deliberately — submit the form empty, trigger the error, load the empty list — and verify each announces itself rather than only changing colour.
9. Re-run the whole pass after fixes. A fix is verified by repeating step 3 onward, never by the reasoning that produced it. Regressions from accessibility fixes are common, especially in focus handling.
10. Stop and ask before anything outward-facing. Publishing an accessibility statement, filling a VPAT or answering a customer conformance questionnaire is a formal representation about the product. This procedure supplies evidence for that; it does not authorize making it.

Checklist

• Scope written down: routes, states, viewport, browser
• The running app was exercised, not the source read
• Automated scan run; tool, version and exact failures recorded
• Full keyboard pass done with no pointer use; tab order written down
• Focus visible at every stop and not obscured by overlaying chrome
• Screen-reader pass on a named reader + browser pairing, completing the real task
• Names, roles and states confirmed as announced — not assumed from the markup
• 200% zoom and 320px reflow checked; text-spacing condition applied
• Contrast measured on rendered pixels in every interactive state
• Error and status messages heard, not just seen
• Everything out of scope or unchecked is named in the report

Failure handling

• No screen reader available — report that the screen-reader checks were not performed, list what was. The claim then stops at "keyboard-operable and scan-clean on ". Do not substitute the accessibility tree for a screen-reader pass: the tree shows what is exposed, not what is announced, in what order, or whether it makes sense.
• The scan reports violations in third-party code — record them separately with their source. They are still barriers for users; they are just not fixable in this diff.
• A barrier will not reproduce — say so with what you tried. Screen-reader behaviour varies by reader, browser and version; name the exact combination.
• Scan is clean but the keyboard pass fails — the keyboard result wins. Report the scan as uninformative here, not as a pass.
• Time ran out mid-pass — report what was covered and what was not. A partial pass honestly scoped is useful; a partial pass reported as complete is worse than none.

What this refuses to conclude

• Zero automated violations is not "accessible" and not "WCAG 2.2 AA".
• A passing keyboard pass is not a passing screen-reader pass.
• One route verified is not the product verified.
• Code that was written following accessible patterns is not verified code — it is unverified code that will probably pass.
• Conformance is a claim about a defined scope against defined criteria. Without the scope and the per-criterion evidence, this procedure produces findings, not conformance.

Evidence to report

Scope; the scan tool, version and output; the tab order as observed, with any traps or invisible stops; the reader + browser pairing and what it actually announced at each step, quoted; the zoom, reflow and text-spacing results; contrast ratios measured, with the states they were measured in; screenshots where a picture carries what words cannot. Then the list of criteria, routes and states not checked. "Accessibility verified" without those is an assertion, not evidence.