2026-07-22
Measured-verdict layer and accuracy benchmark
- NewAdded a measured-verdict layer — target size (WCAG 2.5.8 / 2.5.5), reflow (1.4.10), text spacing (1.4.12), and focus visibility (2.4.7) are now decided from actual pixel and box measurements before the AI is ever asked. The measured values are kept in the evidence for each verdict.
- NewIntroduced an accuracy benchmark — detection and false-positive rates are continuously measured against test pages with known, seeded defects. The deterministic and measured layers must hold 100% detection with zero false positives as a release condition.
- NewJIS X 8341-3:2016 test-result summary — audit results can be summarized (in Japanese) with JIS-covered criteria (the 60 from WCAG 2.0) separated from reference criteria (the 26 added in WCAG 2.1/2.2).
- NewStronger CI integration — SARIF 2.1.0 output (GitHub Code Scanning compatible) and a regression gate that diffs against the previous run to block only new violations, both added to the CLI.
- ModelsAdded support for Gemini 3.6 Flash and Gemini 3.5 Flash Lite (the day after Google's announcement). The existing 3.5 Flash and 3.1-series models remain available.