Technology
Two systems. One writes fixes, the other proves them.
The architecture exists to make one guarantee: no fix is marked verified by the system that produced it. Everything else follows from that.
Crawler
An edge crawler that sees what users see
Discovers templates, components and interactive states from the sitemap, the router and scripted navigation. Runs close to the site, respects robots rules and rate limits, and captures the rendered DOM and accessibility tree rather than raw HTML.
Fixer
Constrained drafting, reviewed by engineers
Language models draft code changes inside a rule set we wrote per defect class: what may change, what must not, and which criterion the change must satisfy. Drafts are reviewed by an engineer before they become a pull request.
Verifier
Independent checks, three ways
Automated rules on the rendered page; recorded runs with NVDA, JAWS and VoiceOver in a controlled browser farm; and keyboard-only runs. Deterministic checks calculate contrast from browser-derived colours. Image analysis, run as batched GPU inference, supports complex backgrounds and flags image alternatives for contextual human review. The verifier has no access to the fixer's reasoning, only to the result.
Evidence store
Every claim linked to its proof
Each criterion result stores the method, the artefact (recording, data table or annotated screenshot), the reviewer decision where one was needed, and the commit it was tested against. Reports are generated from the store, so they can be reissued for any date.
Security and data
Nothing runs on your visitors' browsers.
- No script tag, overlay or runtime agent on your site; the product works on your code and on crawls of your pages
- Read-only repository access is enough to start; write access is optional and limited to opening pull requests
- Test accounts, recordings and crawl data are deleted on request, and recordings never include real customer data when a staging environment is provided
- Crawls and verification run in our cloud environment in the EU or the US, at your choice
Open source
The rule layer will be public.
We plan to publish the automated rule layer as an open-source command-line tool, so anyone can run the same first pass we do, free. The assistive-technology runs and the verification engine are the product. Until then, you can run automated rules on your own HTML in the browser.
See it on your own product.
Start with one agreed user journey at no cost through early 2027. We agree the defects, deliverables and your review effort before the pilot begins.