A Worked Section 508 Evaluation

What Happened

Automated analysis of census.gov flagged issues across 8 success criteria. Verifying every one by hand against the live page on July 31, 2026 produced 10 confirmed defects and threw away 15 findings in four classes, each listed below with its count and the reason.

That ratio is the reason a conformance report needs a person attached to it. Every eliminated item below was plausible, specific, and wrong, and every one would have reached a client as a defect to fix if nobody had checked. Six claimed that hovering revealed content when the handler actually dismisses it.

Verification cut the other way too, and we are showing that rather than quietly fixing it. Two defects we dismissed on the first pass turned out to be real, and both were dismissed for the same reason: the check we ran could not see the thing it was looking for. One heading-structure defect lives inside an embedded frame our heading walk never entered. One sits inside an element we had waved away on a technicality about its tag. A third finding we called a false positive - an undersized tap target - was correct all along; we had applied the spacing exception with the wrong test.

Scope, stated precisely. One page, the homepage, at 1440x900, in Chromium, on July 31, 2026. We verified the failures, not the passes, so nothing here should be read as a clean bill of health for the criteria not listed. A single page sample can show that a site does not conform; it can never show that it does. This is a methodology demonstration on a public site, published unsolicited, with every item reproducible from the evidence given.

The 15 Findings We Threw Away

Grouped into the four claims they came from, with exact counts so the arithmetic reconciles. This table is the actual work product of a manual audit, and it is the part an automated report cannot produce. Every class below has since been corrected in the tooling - see the section after it.

What We Did About It Afterwards

Throwing away a bad finding helps one client. Fixing the logic that produced it helps every client after them. So each class of false positive above was traced back to the code and corrected.

The hover check graded findings on the mere presence of an onmouseenter attribute and never read the handler; it now reads the handler body, drops the ones that only dismiss, and hedges when it genuinely cannot tell. The obsolete-element rule no longer reports presentational tags under a criterion about interactive components. The missing-role check no longer flags a wrapper whose focusable child already exposes a role. The javascript: URL rule stopped asserting a keyboard failure it had no evidence for.

Re-running the same scan afterwards: 9 findings removed, none added, and every confirmed defect on this page still fires, including the two the first verification pass had wrongly dismissed. That last part is the test that matters. A fix that quietly silences a real defect is worse than the false positive it replaced, so the check is always run in both directions.

A Note on the Site We Tested

census.gov was chosen because it is public, high-traffic, and squarely within Section 508’s scope, not because it is unusually poor. Several things on it work well: the skip link is the first tab stop, becomes visible on focus, and moves the reading position past the masthead, which is more than many commercial sites manage. The findings above are offered constructively and every one includes the fix. If anyone at the Census Bureau wants the underlying evidence, it is available on request.