# Licensing and provenance policy

UIJudgeBench is a hybrid corpus. Every page carries a `provenance.json` recording its
source, license, and retrieval/generation date, and every label line records the same.
This document is the authoritative per-source policy. When a source's terms are
unclear, we investigate and document rather than guess; where redistribution is not
clearly permitted we ship the download script instead of the data (the "LAION model").

## Code

- **This repository's code** — MIT (see `LICENSE`).
- **Vendored axe-core 4.10.3** (`uijudge/vendor/assets/axe.min.js`) — Mozilla Public
  License 2.0; full text in `uijudge/vendor/assets/LICENSE-axe.txt`. Copied unmodified
  from the LayoutLens project (MIT). The browser/axe wrappers in `uijudge/vendor/` are
  derived from LayoutLens (`layoutlens/browser.py`, `layoutlens/a11y/`); see
  `uijudge/vendor/NOTICE.md` for attribution.

## Corpus sources

### W3C ACT test cases  (door: `ingested`)

- **Manifest:** `https://www.w3.org/WAI/content-assets/wcag-act-rules/testcases.json`
  (verified live; 1,188 cases across 87 rules; self-declared license link
  `https://act-rules.github.io/pages/license/`).
- **License (verified):** the ACT-Rules Community Group publishes the rules and their
  test cases under the **W3C Software and Document Notice and License**. This is the
  *combined* W3C license (not the modification-prohibited "W3C Document License"): it
  explicitly permits copying, modification, and redistribution *with or without
  modification*, for any purpose, provided the NOTICE and a copyright/attribution
  statement are preserved.

  > Correction to the P1 brief: the brief described these as "W3C Document License —
  > modification not permitted." The actual license (verified at the URL above on the
  > retrieval date recorded in each `provenance.json`) is the more permissive W3C
  > Software **and Document** license, which *does* permit modification. We nonetheless
  > **redistribute the test-case HTML byte-for-byte unmodified** — a deliberate choice
  > for provenance integrity and reproducibility, not a license requirement.
- **Canary policy:** because we redistribute these HTML files unmodified, the canary is
  **NOT** injected into the ACT HTML. It lives only in our sidecar `provenance.json`
  and in the `labels/items.jsonl` entries.
- **Attribution:** each page's `provenance.json` records the source URL, license name,
  license URL, and retrieval date. The required W3C NOTICE is satisfied by the license
  reference recorded per page plus this document.
- **What we ingest:** `passed` and `failed` cases that map to at least one primary WCAG
  success criterion. `inapplicable` cases are skipped for v1 (no yes/no page verdict).

### GDS accessibility-tool-audit  (door: `ingested`)

- **Source:** `https://github.com/alphagov/accessibility-tool-audit`
  (`tests.json`, 142 seeded-barrier cases across 19 categories).
- **License (verified):** **MIT** (repo `LICENCE` file). Redistribution and modification
  permitted with attribution.
- **Canary policy:** these pages are *generated by us* (we wrap each documented barrier
  snippet in a minimal HTML document), so the canary **IS** embedded as an HTML comment.
- **What we ingest:** every case is a documented accessibility barrier (a violation by
  construction), so each yields an L1 page verdict (`no` — page does not satisfy the
  category) and an L2 defect-typing item. Criterion codes use the `gds:<category>`
  namespace rather than a fabricated WCAG SC, because upstream labels barriers by
  category, not by success criterion.

### AccessGuru  (door: `ingested`, download-at-build)

- **DOI:** `10.18419/DARUS-5177` → DaRUS (Universität Stuttgart Dataverse).
- **License (verified):** **CC BY 4.0** (`http://creativecommons.org/licenses/by/4.0`),
  confirmed via the DaRUS dataset API. Redistribution and adaptation permitted with
  attribution — so we *may* redistribute.
- **Decision:** we ship the **download-at-build ingestion script**, not the data. The
  dataset is ~552 MB (a 494 MB screenshot archive dominates); committing it would bloat
  the git repo with no benefit for a P1 walking skeleton. The tabular file
  `Original_full_data_new` (~20 MB, 3,524 rows) is the useful part; the ingestion module
  fetches it on demand via the DaRUS access API. This is a size/hygiene decision, not a
  license one — CC BY 4.0 would permit in-repo redistribution with attribution.
- **Attribution (required by CC BY 4.0):** Fathallah, N., Hernández, D., & Staab, S.,
  *"AccessGuru: Leveraging LLMs to Detect and Correct Web Accessibility Violations in
  HTML Code"*, DaRUS, DOI 10.18419/DARUS-5177, CC BY 4.0.
- **Status (P3):** license verified; the download-at-build module
  (`uijudge/engine/ingest/accessguru.py`) maps a curated slice of 133 violations
  (Layout 50, Semantic 33, Syntax 50, across 62 pages) into `ingested`-door L1 items.
  Because the tabular data gives the affected element only as an HTML fragment (no CSS
  selector, no bbox), items map to `annotation_unit=page` with the element fragment + axe
  rule + taxonomy class preserved in the receipt (see `docs/UNITS.md`). Attribution is
  recorded in every item's provenance. Bulk data (screenshots + page archive) is never
  committed; the ~20 MB tabular file lands under `corpus/_downloads/` (git-ignored).
- **Quarantined in v0.1.0 (composition, not license):** the mapped slice is **held out of the
  scored corpus** because its pages are not materialized (no `page.html`) and every ground
  truth is `"no"`. The ingest emits to `labels/quarantined/accessguru_items.jsonl`; the CC BY
  4.0 grant above still governs the data. See `labels/quarantined/README.md` and `datasheet.md`
  Known limitations #11.

## Real-page policy (P3)

Real web pages enter the corpus under a strict two-tier policy driven by
`uijudge/engine/real_manifest_v1.json`.

### Tier-A — redistributable, committed

Frozen into self-contained artifacts (`corpus/real/<page_id>/`) and committed, with the
canary injected and stable `uij-e*` element ids assigned (see `docs/UNITS.md`). Sources:

- **US federal `.gov` pages** — public domain under **17 U.S.C. § 105** (works of the U.S.
  federal government are not subject to copyright). Text and layout are PD; embedded
  third-party assets are excepted, and our inlining caps bound what is mirrored while
  provenance records the source URL. This is the bulk of the roster (~50 pages spanning
  usa.gov, cdc.gov, nih.gov / medlineplus.gov, nasa.gov, ada.gov, section508.gov,
  digital.gov, and others), tagged across genres (landing, gov-service, docs, blog).
- **Explicitly permissive OSS docs** — SQLite docs (**public domain**), Python docs
  (**PSF License Agreement**, reproduction permitted with notice), Django docs
  (**BSD-3-Clause**). Each carries its specific license URL + evidence quote in the manifest.

Every included URL records its license, license URL, and an evidence string. Per-page
`provenance.json` and every emitted item's provenance repeat the license.

### Tier-B — ship script, not content

Proprietary pages (e.g. github.com, nytimes.com, airbnb.com marketing pages) with no
redistribution grant. The freeze CLI writes them to **`corpus/real/tier_b/`**, which is
**git-ignored** (`.gitignore`), so their content never lands in the repo — we ship only the
URL + the deterministic freeze script (the "LAION model"). The manifest carries 2–3 tier-B
examples to exercise this path; a test asserts `git check-ignore` excludes the path.

### Rejected candidates (honesty artifact)

Recorded in the manifest's `rejected` list with a reason each. We declined:

- **MDN** and **Wikipedia** (CC-BY-**SA**) and **Stack Overflow** (CC-BY-SA) — the
  share-alike clause would force our snapshot/corpus under copyleft, conflicting with the
  corpus's uniform CC-BY/MIT licensing. Not forbidden reuse; a deliberate
  copyleft-entanglement avoidance.
- **Smashing Magazine** — no blanket redistribution license found (all-rights-reserved by
  default); license uncertain, so excluded rather than guessed.
- **Bootstrap docs** — code is MIT but the docs content is CC-BY-3.0 with a separate
  attribution requirement; excluded to keep corpus licensing clean.

Synthetic corpus is CC-BY-4.0. Provenance is recorded per page and per item.