Accessibility tooling · Product engineering
Accessibility Evidence Engine
A repeatable system for making accessibility claims from evidence instead of scores.
Automated accessibility tools are useful, but they answer only part of the question. I built the Accessibility Evidence Engine to keep automated checks, deterministic browser evidence, expert review, genuine human acceptance, remediation and retesting separate — and to make unfinished evidence stay visibly unfinished.
Active internal system · WCAG 2.2 A/AA-oriented evidence workflow
The problem
A clean automated scan is not an accessibility conclusion.
Automated tools can find real failures quickly, but they cannot establish whether every relevant interaction works for a person, whether a screen-reader experience makes sense, or whether all applicable WCAG requirements have been satisfied.
The failure mode I wanted to remove was evidence collapse: an axe result becomes an “marketing metric”, a successful code change becomes “fixed”, or several passing checks become a blanket WCAG claim.
Missing evidence should stay missing. It should not quietly become PASS.
The workflow
Test. Investigate. Fix. Retest. Keep the boundary visible.
- ScopeDefine the pages, interactions, criteria and evidence required before testing begins.
- Automated checksUse automated accessibility testing for the failures it can reliably detect, without treating a clean scan as proof of conformance.
- Deterministic browser evidenceCapture repeatable evidence for structure and behaviour such as headings, landmarks, focus, text spacing and reflow.
- Human acceptanceKeep checks that need genuine human judgement — including screen-reader acceptance and other bounded manual gates — explicitly open until they are actually tested.
- RemediateTurn evidenced failures into bounded implementation fixes rather than broad speculative rewrites.
- RetestRun the same gates again after remediation so a fix is accepted from evidence, not from the fact that code changed.
- Publish the boundaryRecord what passed, what still needs human acceptance and what cannot yet be claimed.
Evidence model
Different evidence answers different questions.
Automated evidence
Playwright and axe provide repeatable automated coverage for issues they can detect reliably. Violations are retained as evidence and traced through remediation and retest rather than reduced to a percentage score.
Deterministic browser evidence
Browser and accessibility-tree checks capture structural and behavioural facts that can be tested consistently: heading order, landmarks, accessible names, focus behaviour, text-spacing resilience, reflow and related implementation gates.
Expert and human evidence
Checks that require genuine human judgement remain separate. Screen-reader acceptance, native zoom behaviour and other bounded manual gates do not inherit PASS from automated evidence.
Durable evidence packs
Results are kept as durable evidence rather than disappearing into a chat or terminal session. That makes it possible to distinguish the baseline, the failure that was investigated, the remediation, the retest and the final accepted boundary.
Claim semantics
- PASS
- FAIL
- MANUAL_ACCEPTANCE_REQUIRED
- NOT_READY
- EVIDENCE_COMPLETE
EVIDENCE_COMPLETE is a bounded evidence state, not a universal certification. The engine does not produce an accessibility percentage and does not turn partial evidence into a blanket WCAG conformance claim.
Used on real work
The system has to survive contact with Production.
Anders Norrman
The Anders Norrman Production remediation completed its defined evidence scope: 7/7 sampled pages were audited, the final Production axe run contained zero violations, identified failures were fixed and retested, and the bounded project evidence reached EVIDENCE_COMPLETE.
That status describes the accepted audit scope. It is not a blanket WCAG-conformance claim for every possible page, state or future change.
Bröd by the Bay
Bröd by the Bay demonstrates the other side of the model: automated and deterministic remediation evidence can be strong while genuine screen-reader acceptance remains a separate open gate. The system keeps that difference visible rather than promoting the project to an unsupported completion claim.
The result
Accessibility evidence that is useful precisely because it refuses to overclaim.
The engine turns accessibility work into a reproducible product workflow: establish a bounded scope, collect the right kind of evidence, remediate what failed, retest the same gates and preserve whatever still requires human acceptance.
Accessibility & product engineering
Need accessibility work that ends with evidence, not a score?
I can scope, investigate and remediate accessibility barriers while keeping automated evidence, manual acceptance and claim boundaries explicit.
Discuss accessibility work