# OnCo reference-slice conformance test (local opt-in)

Purpose: validate content-explorer against the *real* upstream UX and data topology, without treating the OnCo corpus as product data. This is an additional test track, not a replacement for CC0 generic fixture tests.

## Non-negotiable rights boundary

- OnCo **software** is MIT; OnCo **curated dataset** is CC BY-NC 4.0, with its own `LICENSE-DATA`. This repository contains **no copied OnCo dataset**.
- The upstream data terms expressly classify **a company's internal data pipeline** as commercial use requiring the rightsholder's agreement. **Small slices and private tests do not automatically create an exception**. Before importing real records in an organizational/commercial workflow, obtain written permission or a suitable license, or get independent legal approval for a defensible exception. Until then, use synthetic fixtures and references/URLs only.
- Record authorization in a *private* `authorization.json` document. Do not commit it if it contains contacts/confidential terms. Publication and CI distribution require separate permission evaluation.
- Attribution wherever permitted OnCo content is presented: `Data from OnCo (onco.cc)` linking to https://onco.cc; license link https://creativecommons.org/licenses/by-nc/4.0/; source URL, upstream SHA, accessed date, modification notice, and original upstream links. Inspect third-party assets separately; no logo/portrait copying by default.
- No patient/personal health records, no clinical conclusions, no medical-advice claims.

## Reference slice selection (target ~12–25 nodes, not a corpus fork)

Choose a connected induced/test subgraph from an immutable upstream commit. Include one of each: cancer, treatment/drug, target, technology, clinical trial, key paper, institution and glossary term; optionally a person and idea if rights permit. Choose real items **after** authorization and upstream pinning, not by hard-coded contemporary medical claims.

Coverage requirements:
1. At least two different types of edges and an inverse/backlink traversal.
2. At least one one-to-many relationship and one multi-hop path (depth >= 3).
3. At least one entity with >1 provenance link; dated TL;DR and technical summary.
4. At least one missing optional field, and at least one unresolved relation handled explicitly (not silently erased).
5. Two filter facets, full-text query and a stable URL deep link.
6. A few controlled mutations: renamed title, removed node, bad cross-reference, duplicate id, license metadata removed (all modifications clearly tagged as local tests, never represented as upstream facts).

## Proposed workflow

**Gate A — authorization:** verify rights and record intended use; CI containing the real dataset must remain disabled without sufficient grant.

**Gate B — pin & extract:** pin upstream commit (`git rev-parse HEAD`), document source paths and entity ids in `slice-selection.yaml`; extractor reads only enumerated records. Compute sha256 of originals and emitted projection. No bulk download/auto-discovery as a side effect of normal builds.

**Gate C — dual oracle:** run the pinned OnCo implementation as reference and capture results for identical user tasks, using stable screenshots and structured browser/graph assertions. Use the *same pinned records* through the connector to make comparable content states; preserve expected divergences in an explicit exception list.

**Gate D — conformance:** test source -> projection -> view model -> route using schema tests, semantic invariants, Playwright browser journeys, and visual snapshots. Record each mismatch as extraction, mapping, semantics, presentation, or upstream instability.

**Gate E — reproducibility:** publish hashes, commit SHA, normalized test descriptions and aggregate results only when permitted; distribute synthetic fixtures as the default executable public tests. Never upload the OnCo slice as a CI artifact without authorization.

## Run profiles

- `public-ci`: synthetic fixtures only; completely offline and distributable.
- `licensed-reference`: locally provisioned real slice, explicitly enabled after authorization; never default, no automatic telemetry/artifact export.
- `visual-upstream`: uses browser captures from pinned upstream locally; screenshots may have independent IP considerations; approval needed before publication.

## Acceptance gates

- [x] Local authorization and permitted scope recorded before extraction. Redistribution and public CI stay refused. The private file is gitignored.
- [x] Slice allowlist enumerates the pinned upstream SHA, 17 ids, files, and types. Provenance and the CC BY-NC notice are applied by the adapter.
- [x] Extractor is allowlist-only. Edges outside the allowlist are reported and are not stored.
- [x] Projection keeps identity, types, in-slice relations, as-of metadata, missing optional fields, and the rights notice.
- [x] The licensed test requires zero silent link loss and zero attribution gaps. Public behavioral tests use synthetic records.
- [x] Visual pixel parity is an exception: the upstream UI is not rebuilt here, and slice screenshots are not published. Semantic and HTTP checks are in `workplans/evidence/CONTENT-WP-0007-Review.md`.
- [x] Public CI passes with no OnCo corpus. The licensed catalog root is off unless `CONTENT_EXPLORER_LICENSED=1`.
- [x] The slice and the synthetic dataspaces use the same UI components. Kind labels come from the manifest.

## Evidence

`reference-manifest.json`, `baseline-report.md`, `projection-diff.json`, `journey-results.json`, `visual-review.md`, `rights-review.md` (non-confidential). Keep raw licensed data in ignored local directories.