Test report

Recorded 2026-11-21T20:13:16.345Z on Node v24.15.0, win32/x64. Version 0.2.1. Generated by npm run proof; nothing here is hand-edited.

Results

MeasureResult
Unit, integration and security tests74 run, 73 passed, 0 failed, 1 skipped
Assurance checks20 of 21 passed, 0 failed, 1 skipped
VerdictASSURANCE_CHECKS_PASSED_HUMAN_REVIEW_STILL_REQUIRED
Gold corpus56 cases; precision 100.0%; recall 100.0%
Fuzzing2213 inputs; 0 unexpected exceptions; slowest 253 ms

The skipped test and check are the symlink-refusal tests, which cannot create symlinks without privilege on Windows. They run on Linux and macOS CI.

Real browser cross-check

Chromium 131 with Playwright 1.49.1 and @axe-core/playwright 4.10.0 on both fixtures: the broken fixture produced axe violations button-name and image-alt (one node each); the good fixture produced none; 27 axe rules passed on each. The static scanner reported the same two plus the placeholder-only field and a clickable div for review. See test-results/axe-crosscheck.json and the raw browser-*.json.

Bugs found by this process and fixed in 0.2.0

A multi-minute parser hang on 2 MiB of < (found by fuzzing); Windows installer containment bug; the CLI doing nothing under an npm bin symlink; select/textarea false negatives; implicit-close nesting false errors. Full list in the changelog.

How accurate is "100%"?

The corpus was written by the maintainer from HTML and ARIA semantics before looking at tool output, then used to check the tool. It is small and not independent. Treat 100% as "consistent with its own spec", not as a real-world accuracy figure. Independent labelled data is on the roadmap.

Raw artifacts

test-results/test-run.tap, benchmark.json, fuzz-results.json, environment.json, SHA256SUMS.txt; transcripts under samples/; sample assurance report in samples/audit/.