CurbCut: Accessibility Review That Proves Fixes

AI coding assistants have become remarkably capable, and more and more interface code is now written with their help. Their output, however, still tends to overlook the accessibility standards that people with disabilities depend on: buttons without names, text with too little contrast, forms that cannot be completed with a keyboard. Speed has improved; inclusion has not kept pace. Accessibility issues pass code review because most reviewers are not WCAG experts, and linters only catch what a rule can match. Students and citizens who use screen readers or keyboards find the gap after release. CurbCut puts an accessibility reviewer into every pull request, built entirely from IBM Bob's own building blocks. Bob read the WCAG 2.2 Recommendation and wrote a reusable skill with a policy pack of all 55 Level A and AA criteria. In the CurbCut Review mode, Bob runs a deterministic engine for measured facts, then spawns four read-only Explore subagents in parallel, one per WCAG principle. Every finding is labelled proven, backed by a measured fact, or flagged, Bob's judgment with a reason, and checks that cannot be done this way are always listed. The CurbCut Fixer mode can only edit the app and makes the smallest fix per finding. The CurbCut Prover mode writes one test per fix, and a fix counts only if its test fails on the pull request and passes on the fix. On a demo class app with planted issues, axe-core alone found 8 of 18 issues in the first pull request. CurbCut with Bob found 16, flagged no decoys, and proved 23 of 23 fixes. Results are shown on a live dashboard with a screen reader narration before and after, and exported to GitHub code scanning. CurbCut does not certify conformance, and it is no substitute for testing with disabled people.