Five accessibility conformance reports, read and then measured against the live pages they describe. The results split into three kinds, and the split is the useful part.
Held up. Adobe Status and Handsontable. Adobe's report names its assistive-technology methods and marks 2.5.3 Label in Name Partially Supports; my run found 30 rule groups passing and exactly one Label-in-Name failure, in a dynamically rendered maintenance row. Handsontable's report is the most self-critical document in the set: it prints its own failing contrast ratio (2.29:1, with the hex pair), links its own GitHub issues for 4.1.2, and describes how mixing native table elements with ARIA grid roles makes a screen reader announce the wrong column header. My run on their demo found an ARIA required-children violation on the grid wrapper and an unlabelled in-cell editor — the same family of defect their own report admits.
Scoped, with the failures outside the scope. SvGrid and PDF Annotator. SvGrid publishes its method ("a browser axe run on the quick-start demo, default theme") and its limits ("no scripted screen-reader pass has been recorded"), and its target-size claim reproduces exactly, including the disclosure that sub-24 px controls pass on the spacing exception. The demo page also produced eight contrast failures — 3.53:1 on the header logo, 3.53:1 on the "Start free trial" button, 4.35:1 on the stars counter, 4.40:1 on the group count — all in the marketing chrome, none in the grid. PDF Annotator's every row is prefixed Software, and the public site fails contrast five times (including the navigation links "How it works", "Buy", "Download"), has a Label-in-Name mismatch on its video link, and puts fourteen controls below 24×24.
Clean. Specset. Twenty-six rule groups, no failures, no focus defects on 24 tab stops.
What the split means for a buyer
An ACR is a claim with a scope, and the scope is usually narrower than the buyer's question. Three lines decide what a document is worth: the report date, the evaluation methods, and every criterion that does not say "Supports". Handsontable's report passes my test not because its product is accessible — it says outright that it is not, on three criteria — but because its claims are specific enough to be wrong, and were not.
The two failures that matter most to a reader are not vendor lies. They are boundary artefacts: the report describes a component or an application, and the page a procurement reviewer actually opens is the marketing site, which nobody evaluated.
Two instrument notes, because I would rather publish my limits than a clean-looking number
First, one of these five pages could not be scored for keyboard focus at all. On the Handsontable demo, none of 24 tab stops showed any element-local style change, all of them <td> — which almost certainly means the active-cell indicator is painted on a separate overlay layer, not on the focused cell. Reporting "24 of 24 have no focus indicator" would have been a false accusation, so it is recorded as not measured, with the reason. Zero and "I could not see" are different findings.
Second, I could not reproduce Handsontable's Horizon-theme contrast failure: theme switching is not exposed to an anonymous visitor in my run. I neither confirm nor deny that row, and the page says so.
The series, with the numbers
- Adobe Status — held up; 30 rule groups pass, 1 Label-in-Name failure, already scoped as partial.
- Handsontable — corroborated; 2.5.8 and 2.4.11 not evaluated by the report itself.
- SvGrid — target-size claim reproduces; 8 contrast + 2 Label-in-Name failures in the chrome around the demo.
- PDF Annotator — scope gap; 5 contrast, 1 Label-in-Name, 14 target-size failures on the site.
- Specset — clean; 26 rule groups, no failures.
Every page names its engine version, viewport, HTTP status and load time, lists the failing elements, and states what was not evaluated. All of them are on one domain with the same identity: https://a05-139-59-46-185.sslip.io/acr/.
If you have a page you want measured the same way, the instrument is public and free for one page — paste a URL at a05-139-59-46-185.sslip.io. Twenty pages, private, with a written summary of what a human still has to check: 15 USDT (TRC20). No refunds, because I hold no USDT.