Open Accessibility Workbench
Turn a wall of accessibility findings into a small, role-aware set of explainable remediation tasks, then give each person the evidence, context, guidance, and validation they need to take the next useful action.
🚀 Key Features
Status legend: ✅ completed · 🟡 scaffolded (degrades to a deterministic path) · ⬜ planned. See docs/ROADMAP.md for the authoritative breakdown.
- ✅ Provenance-Preserving Reduction: Transparently reduces thousands of scanner observations down to a handful of actionable tasks without silently discarding findings.
- ✅ Multi-Scanner Ingestion: Direct support for Open Scans (
report.json, report-overlap.json, report.csv) and Oobee (report.csv, summary JSONs, .json.gz.b64). A summary report.csv loaded by URL auto-loads its finding-level report.json sibling.
- ✅ Deterministic Pattern & Component Engine: Correlates cross-page structural DOM signatures and upstream pattern hashes to identify probable shared components (e.g. social link clusters, site headers, footers).
- ✅ Multidisciplinary Role Routing: Leverages W3C ARRM (Accessibility Roles and Responsibilities Mapping) and user capability profiles to route tasks across Content, Design, Frontend, QA, and Product roles.
- ✅ Deterministic Remediation Blueprints: Generates step-by-step blueprints, human decision checklists, and native HTML remediation guidance without requiring AI.
- ✅ Client-Side Privacy Guarantee: 100% in-browser processing. Zero data uploads, zero telemetry, no cloud AI API keys required. Works offline after one online visit; large reports are analysed in a Web Worker with progress and cancel.
- ✅ Multi-Format Export: Full provenance JSON, JSON-LD, Markdown, and GitHub issue handoffs.
- 🟡 Optional In-Browser AI with Validation Loop: The consent gate, untrusted-data-safe response processing, invention checks, bounded validation loop, and a real on-device model runtime (transformers.js with download/cancel/dispose, WebGPU/WASM inference, a Hugging Face or GitHub-release weights source, and AI output shown as a clearly-labelled draft) are all built. The runtime is build-gated (
VITE_AI_RUNTIME=1) and OFF in the default build, which downloads no model and tree-shakes the ~100 MB dependency out. It stays scaffolded until verified on WebGPU hardware — see the roadmap. Inference is on-device; your report is never uploaded.
🛠️ Quick Start
# Clone the repository
git clone https://github.com/mgifford/open-accessibility-workbench.git
cd open-accessibility-workbench
npm install
npm run dev # local dev server
npm test # unit tests (verifies bundled data provenance first)
npm run test:browser # Playwright browser + accessibility tests
npm run build # production build to dist/
npm run build:data # regenerate bundled ARRM / guidance / RAG data
The live build is deployed to GitHub Pages on every push to main, gated on both
the unit and browser test suites.
♿ Accessibility & Privacy
- Accessibility Statement — WCAG 2.2 AA target, testing
performed, known limitations, and how to report a barrier. We do not claim
conformance on the strength of automated tests alone.
- Privacy Policy — local-only processing; what may be stored on your
device (opt-in preferences only, never report contents).
- Roadmap & Capability Status — what is completed, scaffolded,
and planned.
📖 Documentation
📄 License
GPL-3.0-or-later. See LICENSE for details.
