Accessibility

What has been tested, what it measured, and what has not been tested at all.

Target conformance level: WCAG 2.1 Level AA. This page is not a VPAT. It is the evidence a VPAT would be written from, published while the product is small enough that you can check it yourself.

The last measurement

Run against the live site, in Chrome, at a Chromebook viewport and a phone viewport, on Sep 1, 2026, 2:03 AM UTC. Re-run on every deploy that changes a page.

The routes checked were: /, /why, /for-educators, /compliance, /privacy, /explore, /record, /accessibility, /work, /work/oh, /talk, /404. A build check fails if a route ships that is not on this list. One of those, /work/[state], is a template that renders a page for every state, so about fifty pages share its code and /work/oh is the one sampled. That is a real limit of this measurement rather than a technicality: a defect present only on a longer state name would not appear in the number above.

Automated WCAG 2.1 A and AA violations: 0Checked on all 12 routes at 3 viewports (1366x768, 390x844 and 320x800) in BOTH the light and dark palettes, using axe-core restricted to the wcag2a, wcag2aa, wcag21a and wcag21aa tags. The completed results screen — eighteen occupations, the filter, the state picker — is scanned separately at both viewports and is included.
Routes with horizontal scrolling: 0 of 12Measured as scrollWidth minus clientWidth on the document element, at all three viewports. Note what this cannot see: content clipped by overflow:hidden does not overflow, so this figure is not a reflow test. A separate check of truncated text found and fixed nine clipped labels at 320px that this number reported as clean.
JavaScript console errors: 6Across 72 page loads. An error here usually means something did not render, which no accessibility rule will catch.
Sixty questions completed with the keyboard alone: Yes, in 269 keystrokesTab, Space and arrow keys only, with no pointer at any point, run at BOTH viewports. The run answers a differentiated profile and the pass requires 18 occupations to have actually rendered — not merely that the page changed. An earlier version of this figure ran a straightlined pattern, ended on the screen where the instrument refuses to score, and reported success anyway.
Focused controls not fully visible on screen: 0 of 120Measured on the LABEL, which is the control a person sees, at both viewports. The input itself is visually hidden at 1x1px and is always scrolled into view, so measuring it — as this figure previously did — cannot fail. A control that is focused but not visible is a control a sighted keyboard user has lost.

How it was tested

Chrome, in an isolated browser profile, against the deployed site rather than a local build.
Three viewports: a Chromebook (1366 by 768), a phone (390 by 844), and 320 by 800, which is what 1280px looks like at 400% zoom and is the width SC 1.4.10 Reflow is written about.
Both color schemes. The site ships a full dark palette and an earlier version of this measurement never ran against it.
axe-core, restricted to the WCAG 2.1 A and AA tags, so the number is against the level we claim and not against a broader best-practice set that would flatter it.
A full sixty-question run driven by keyboard events only, at both device viewports, asserting that occupations rendered rather than that the page changed.
Color contrast recomputed from the rendered pixels rather than read from the stylesheet — including a pixel diff of each primary control focused against the same control blurred, which is the check that catches a focus ring the same color as the page behind it.

What has NOT been tested, and what an automated pass cannot see

A clean automated run is a floor, not a result. Two defects on this site passed every automated check and were found by computing a value by hand and by looking at the page.

No screen reader testing. No JAWS, no NVDA, no VoiceOver. The keyboard path and the live-region announcements are verified programmatically, which is not the same as a person using this with speech and telling us it makes sense.
No third-party audit and no VPAT or Accessibility Conformance Report. We will not send you one we have not earned. When there is a customer to justify the engagement, the remediation is budgeted as the expensive half, before the audit rather than after it.
Chromium only. No Firefox and no Safari, and half the context we design for is a phone.
One of the routes above is a template. /work/[state] renders a page per state, so about fifty pages share its code and only /work/oh is sampled. A defect present only on a longer state name would not appear in these numbers.
An automated pass is blind to whether a control is DISABLED or merely says it is. Swapping a disabled attribute for aria-disabled on this site kept the control active, which pulled its text and its boundary out of two contrast exemptions at once. Both then failed, at 3.23:1 and 1.58:1 against requirements of 4.5:1 and 3:1, and axe reported the rules inapplicable rather than failed.
It is blind to a focus ring that is the same color as the page. Every primary control on this site shipped with one — measured 1.00:1 against a 3:1 requirement, with the focused and unfocused renderings byte-identical — under a clean axe run, until an auditor diffed the pixels.
It is blind to meaning. It cannot tell you that a live region announced the same string twice and so said nothing the second time, that a filter replaced eighteen results in silence, or that a discriminator added to defeat that gets SPOKEN as a number. All three shipped here.
It is blind to a class that styles nothing. Six links used a utility class defined in no stylesheet and rendered as plain text, indistinguishable from the sentence around them, which fails SC 1.4.1. A screenshot caught it.
No testing with cognitive or learning disability users, and no plain-language review by anyone outside this project. We measured the results page at roughly grade 11 on Flesch-Kincaid and an independent auditor measured 12.9 on the same text, because syllable heuristics differ between implementations. Take the higher one: it is too high for the page a student has to act on, and the cause is unedited O*NET occupation descriptions.
Voice control and switch access are untested.

Why this is your problem and not only ours

The Department of Justice rule at 28 C.F.R. Part 35, Subpart H requires web content and mobile apps provided by a public entity to conform to WCAG 2.1 Level AA. The compliance dates are 26 April 2027 for public entities serving a population of 50,000 or more and 26 April 2028 for smaller ones.

A vendor’s non-conformance becomes the entity’s own, which is why an evaluator is right to ask for evidence rather than a claim. It is also why we would rather show you a dated measurement with its method attached than a document asserting a conformance level nobody outside this building has checked.

Telling us something is broken

Email david@reddenda.com. Say which page, what you were using, and what happened. A barrier reported by somebody who hit it is worth more than another automated run, and we will tell you what we did about it rather than closing the thread.

Related: what is and is not in place across the rest of the procurement questions, and what happens to student data.