Final tranche 01 · Post-fix result

Diagnostic accuracy corpus

This versioned study measures ZemDomu's default source diagnostics against manually reviewed public-repository findings, targeted seeded defects, and a bounded rendered-DOM comparison with axe-core. The published result is reproducible evidence for this tranche, not a claim of perfect accuracy on every project.

Post-fix precision

407/407

100% in this bounded tranche

Seeded recall

37/37

Reviewed defects at recorded anchors

Public repositories

10

538 selected files

Pending axe disagreements

0

All 12 mapped results dispositioned

Scope and pinned versions

The final run scanned lexicographically selected source files from 10 full-commit-pinned public repositories. It used ZemDomu 1.3.19 plus the checked-in Unreleased accuracy fixes, default rule configuration, and isolated source analysis. Cross-component analysis was disabled for this corpus, so these numbers must not be used as evidence for composed-page behavior.

HTML files
250
React/TSX files
227
Vue files
61
axe-core
4.13.0 · Google Chrome headless

How the precision denominator was built

The natural scan emitted 402 diagnostics: 399 were retained as true positives, 3 Liquid-highlight observations were excluded because they rendered as code examples, and 0 remained unresolved. Another 8 reviewed targeted observations covered default rules without an eligible natural candidate. That produces the published 407/407 precision denominator across all 21 default rules.

Observation syntax counts are 386 HTML, 7 TSX, 14 Vue, and 0 JSX. The absence of a JSX observation is a disclosed sample limitation; it does not establish JSX accuracy.

Precision observations by rule
RuleTrue positivesReviewed denominatorTranche precision
Section heading
requireSectionHeading		2222100%
Heading order
enforceHeadingOrder		1818100%
Single H1
singleH1		66100%
Image alt text
requireAltText		3737100%
Form-control labels
requireLabelForFormControls		4747100%
List nesting
enforceListNesting		11100%
Link text
requireLinkText		11100%
Table captions
requireTableCaption		1010100%
Empty inline tags
preventEmptyInlineTags		11100%
Anchor href
requireHrefOnAnchors		66100%
Button names
requireButtonText		2222100%
Iframe titles
requireIframeTitle		1212100%
Document language
requireHtmlLang		1212100%
Image-input alt text
requireImageInputAlt		22100%
Navigation links
requireNavLinks		11100%
Unique IDs
uniqueIds		11100%
Positive tabindex
noTabindexGreaterThanZero		11100%
ZemDomu placeholders
preventZemdomuPlaceholders		11100%
Document title
requireDocumentTitle		55100%
Single main landmark
requireSingleMain		200200100%
ARIA values
ariaValidAttrValue		11100%

Denominators are intentionally uneven. Nine rules have only one or two reviewed observations, while requireSingleMain has 200. A 100% result with a denominator of one is not comparable to 200/200 and must not be generalized beyond this tranche.

Seeded recall

A seed counted only when its expected rule reported the reviewed defect at the recorded source anchor. Nearby unrelated findings did not count.

accessible-name
10/10
language
10/10
image-alt
10/10
other
7/7

axe-core comparison

axe-core 4.13.0 ran in google chrome headless on three pinned static pages. The mapping recorded 8 agreements, 1 ZemDomu-only result, and 3 axe-only results.

The four differences were documented scope differences: source-level section-heading guidance, a missing-H1 check not present in the measured ZemDomu rule set, and two rendered landmark-containment checks. None remained pending. This is a comparison of selected mapped results, not a claim that the tools are interchangeable.

Pinned repository sample

RepositorySyntaxSelected filesNatural candidatesPinned commit
50projects50dayshtml5094c92b0bc3a87f
mdn-dom-exampleshtml10015672c9e5c6fc61
googlechrome-sampleshtml1001331eef1eeb6048
nextjs-postgres-auth-startertsx70fde8ecf1da93
shadcn-taxonomytsx750298a8857c712
vercel-commercetsx4543761e52e60df
bulletproof-reacttsx10029506629ed003
vuejs-docsvue106b75d188ab16b
antfu-vitessevue1018a01bc9283fe
nuxt-moviesvue416c1abf704762c

Reproduce the evidence checks

From packages/ZemDomu-Core, prepare the pinned public repositories, rerun the reviewed seeds, supply an installed Chrome executable for the rendered comparison, and regenerate the study. Network access is required for the repository preparation step.

npm ci
npm run accuracy-corpus:prepare
npm run accuracy-corpus:seed
npm run accuracy-corpus:axe -- --workspace <prepared-workspace> --chrome <absolute-chrome-path>
npm run accuracy-corpus:study
npm run test:accuracy-corpus

The evidence records the axe-core version but not the exact Chrome build number. That missing browser pin is a reproducibility limitation and should be corrected in a future corpus version rather than inferred here.

Download the raw evidence

Candidate bundle SHA-256
d16f8ac59aa075987ee8b9dfd2f8764a4747d11d063a2f9bee3e6554c2737046
Seed evidence SHA-256
a2f53c97921ab8512e8b892a62f6ec01121f44ff2fcd497da4d49ec130de8084
axe evidence SHA-256
c2595dec2a6e84fcac6db3fe2f16e7ac5981c3df3ad760ab6a21d379594de8ad

For composed-source behavior, inspect the separate cross-component fixture benchmark.