content-provenance.md

# Access Atlas content provenance

This corpus contains 21 public documents: seven guides, seven primary sources, and seven proposed manual checks. It covers button activation, modal focus, basic tabs, actual ARIA menu triggers, keyboard task alternatives, visible focus, and focus obscuration. It was created for Access Atlas on 2 October 2026 and is independent of Deadline Loom's opportunity/deadline records.

The source records preserve the exact URL, live verification timestamp, sections used, informative status, source license, displayed copyright, and a derivation notice. WCAG Understanding records also preserve displayed update dates. [APG evidence](evidence-apg.md) and [WCAG Understanding evidence](evidence-wcag.md) record the primary-source reads. The dataset contains paraphrases and original review prompts, rather than quotations of guidance passages. The W3C copyright notice is retained verbatim.

Each guide references its source and its proposed check. Each check references its guide and source, allowing the agent to separate a repair suggestion from source provenance and a proposed review. All checks have `executionStatus: notRun` and `checkType: manual-review`. Neither these documents nor the golden questions claim an application test, assistive-technology result, conformance audit, legal opinion, or certification.

The APG material is informative pattern guidance. WCAG 2.2 Understanding pages explain individual criteria and are informative companion material, not the normative standard. Several limits are deliberately retained: a role does not implement interaction; modal focus has contextual exceptions; automatic tab activation depends on noticeable delay rather than a numeric threshold; the menu corpus stops at trigger/opening behavior; the keyboard exception concerns a path-dependent underlying function; Focus Visible is separate from contrast and AAA appearance; Focus Not Obscured (Minimum) requires partial component visibility with narrowly scoped notes.

The golden questions specify seven supported repair scenarios and two unsupported questions: contrast calculation and whole-site/legal certification. They include expected guide/source IDs, exact citation URLs, concepts to cover, and claims to avoid. They are authored expectations for an evaluator, not evidence that a model answered correctly. The parent harness records actual model runs separately.

Run `node scripts/validate-dataset.mjs` from the project root to check document types, required fields, unique IDs, reference resolution and reverse relationships, public-source URL allowlist, provenance/license metadata, not-run status, public-data boundaries, and golden citation mappings. This structural validation does not evaluate a rendered interface or certify accessibility.

The current W3C pages link the [2023 Software and Document license](https://www.w3.org/copyright/software-license-2023/). Their content retains those terms. The full notice and source attribution are in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the public [license page](public/licenses.html). Original application code has its separate license; it does not relabel W3C guidance or upstream model/runtime material.