What Automated Accessibility Testing Actually Detects: A Browser-Verified Corpus of 64 WCAG 2.2 Fail and Pass Pairs Evaluated Against axe-core 4.14.0

Preprint. Not peer reviewed. Version 1.0.0. Controlled measurement of what a rule-based accessibility engine actually detects on isolated WCAG 2.2 defects. 64 hand-written HTML fail and pass pairs spanning 36 of the 87 criteria, each rendered in headless Chromium and evaluated against axe-core 4.14.0, with the engine output stored for both variants. The corpus, harness, and raw results are published separately so every figure in the paper can be recomputed. Principal results. axe-core reported at least one violation on 39 of 64 defective documents (60.9%). 63 of 64 corrected documents produced neither a violation nor an inconclusive result (98.4%). One corrected document returned an inconclusive result that still needs human confirmation. Every case predicted as a plain axe violation behaved as predicted (37 of 37). Only one of the three cases predicted to return an inconclusive result produced a violation, so a pipeline that counts violations and drops incompletes records those defects as clean. Across the standard, 29 of 87 criteria (33.3%) have at least one mapped axe-core rule and 58 (66.7%) have none, which bounds what any DOM-based scanner can claim. Level AA, the level adopted by the U.S. DOJ rule and EN 301 549, has the weakest coverage: 7 of 24 criteria (29.2%) have a mapped rule, against 17 of 32 (53.1%) at Level A. Contents. 10 pages, A4. Abstract, background, method, results with four tables, discussion, limitations, reproducibility, conclusion, competing interests, references, and an appendix listing all 36 covered criteria with their detection class and observed result. Files. PDF and the HTML source used to render it. Data availability. Corpus and harness: 10.5281/zenodo.23781867. Criteria reference with rule mapping: 10.5281/zenodo.23781861. Regulatory timeline: 10.5281/zenodo.23781873. Remediation prompts: 10.5281/zenodo.23781879. Reference implementation: github.com/Veduis/wcag-compliance-mcp. Competing interests. The author organisation maintains a commercial accessibility scanning product and a free MCP client built on this corpus. The full corpus, harness, stored output, and classification are published so the measurements can be checked independently, including the cases where the engine failed. Licence. CC BY 4.0. Code in the harness is MIT.