Ir al contenido

Two-tier demo seed dataset (Layer A public / Layer B private)

Esta página aún no está disponible en tu idioma.

Artist Alley needs a realistic, studio-shaped dataset to (a) dogfood federation across two peers, (b) seed local dev + CI, (c) populate the hosted demo once it exists, and (d) produce marketing + README screenshots that show a populated DAM rather than an empty install.

The tension: the most compelling screenshots use rich, recognizable media, but anything shown publicly (README, marketing site, hosted demo, public CI artifacts) must be license-clean — we are publicly displaying every asset in frame. Mixing “attractive but IP-encumbered” content with “safe to publish” content in one dataset makes it impossible to know which surfaces can go public.

The seed dataset is two tiers, both curated under datasets/artist_alley/ (mounted on the self-hosted runner; a curated subset is copied locally for screenshot work):

  • “Artist Alley — Demo Studio Seed (Layer A)” — composed entirely of open-source, public-domain, and Creative Commons content. No IP-referenced or personal content.
  • Aggregate license: CC-BY-SA 4.0 (forced by the most restrictive included source — the Wikimedia gameplay videos are CC-BY-SA; combined with CC0/CC-BY/PD content the aggregate is CC-BY-SA).
  • Ships MANIFEST.json, dataset-metadata.json, and a canonical ATTRIBUTIONS.md listing every source + license. Sources include Kenney.nl (CC0 asset packs), Khronos glTF Sample Models (CC-BY/CC0), Wikimedia gameplay videos (CC-BY-SA), and other CC/PD material.
  • Format coverage spans every asset type the app supports — images, 3D (GLB/OBJ/FBX), audio (OGG/WAV), video (WEBM), documents (EPUB), fonts (TTF), vectors (SVG), HDR — so a seeded instance exercises the full viewer range.
  • This is the ONLY tier used for anything public: marketing-site screenshots, the OSS README screenshots, the hosted demo, and any public CI fixture. It is federation peer “site A” in the dogfood loop.
  • May carry IP-referenced, personal, or otherwise not-publicly- redistributable content that makes for richer internal testing.
  • Never published. Dogfood peer “site B” + local testing only. No screenshots from Layer B reach the README, the marketing site, or the public demo.

CC-BY-SA 4.0 requires attribution + share-alike. Whenever Layer A content (or a screenshot containing it) is redistributed publicly, the ATTRIBUTIONS.md credit list must be preserved / linked. Practically: the marketing site + README screenshot sections carry a “Demo content: CC-BY-SA — see attributions” credit line pointing at the dataset’s attribution list.

Positive:

  • A single, unambiguous rule for “can this go public”: Layer A only. No per-asset licensing audit at screenshot time.
  • Screenshots + hosted demo look genuinely rich (full format coverage) while staying license-clean.
  • The dogfood loop’s two peers (site_a / site_b) map cleanly onto the public/private split.

Negative:

  • Layer A excludes the most recognizable IP, so screenshots skew toward generic-but-clean studio content (Kenney kits, sample 3D models). That is the correct trade for publishability.
  • The attribution credit line is a standing obligation on every public surface that shows Layer A media.
  • Single mixed dataset with per-asset license flags. Rejected — a per-asset audit at every publish point is error-prone; one leaked IP-encumbered thumbnail in a README is a real problem.
  • Fully synthetic / AI-generated demo content we own outright. Considered as a future option; the curated CC/PD dataset ships now and carries real format diversity a generator would have to reproduce.
  • ADR 0049 — Encrypted federation + dogfood (the two-peer loop site_a / site_b feed)
  • ADR 0034 — Capability add-ons (heavy media the dataset exercises)
  • datasets/artist_alley/site_a/{MANIFEST.json, dataset-metadata.json, ATTRIBUTIONS.md} — the canonical dataset provenance

Amendment, 2026-10-05: the dataset no longer has one aggregate license

Section titled “Amendment, 2026-10-05: the dataset no longer has one aggregate license”

The 2026-07-12 decision concluded that Layer A was CC-BY-SA 4.0 in aggregate. That was correct for the sources Layer A held at the time, where CC-BY-SA was the most restrictive term. Layer A has since gained public sources under other terms, including Pexels content under the Pexels License, which is not a Creative Commons license. The current dataset therefore has no single aggregate license. The dataset attribution list was corrected to this model in #1482; this amendment brings the ADR in line with it, and the frontmatter excerpt was updated in the same change.

  • Layer A remains the only public-safe tier, and Layer B remains private. The Layer A and Layer B rules above are unchanged.
  • Public redistribution of Layer A content, or of a screenshot containing it, must honour the terms of every included source.
  • The authoritative licensing and attribution record is the per-asset license and attribution in MANIFEST.json, together with seed/ATTRIBUTIONS.md.
  • Public credit lines describe demo content as coming from mixed-license and public-domain sources and link to the attribution list. They do not name a single license.
  • Earlier wording in this ADR that calls the current dataset CC-BY-SA 4.0 in aggregate is superseded by this amendment. That covers the “Aggregate license” section and the credit line under “Attribution obligation”.

Proposed amendment, 2026-11-11: replace private site_b with a public CI peer

Section titled “Proposed amendment, 2026-11-11: replace private site_b with a public CI peer”

The owner chose to replace the current private site_b with a public-safe corpus for reproducible CI and federation testing under issue #1514. This is a decision about the target corpus, not permission to publish the existing site_b directory. Its Layer B records remain private and must not enter a public repository, release asset, CI artifact, or log.

The replacement will be assembled in a separate staging location from records whose exact bytes have passed per-asset redistribution and attribution review. A record’s layer: "A" label is a candidate filter, not proof of rights. Every current site_b record is excluded from the public replacement; a private per-row ledger records that disposition. Reusing audited site_a bytes is allowed where it helps exercise cross-instance behavior. The two public peers must deliberately share at least two distinct, byte-identical assets as declared fixtures so CI can check the same-file boundary across sites. Those records retain separate peer-local identities.

The replacement becomes the public site_b only after the rights audit, fixture coverage contract, byte verification, and release procedure have passed. Until that cutover, the private site_b rule above remains in force. At cutover, retain the former private corpus only in private storage under a distinct identity; do not copy it into the public release. The site_a rules and the source-specific attribution obligation remain in force.

The release contract and storage choice are separate decisions. This amendment neither chooses GitHub Releases over another host nor changes the existing Kaggle steps in RELEASING.md.

Amendment accepted, 2026-11-15: the public site_b peer

Section titled “Amendment accepted, 2026-11-15: the public site_b peer”

The proposed amendment above is accepted, in the form below. Its text is kept as the record of what was proposed on 2026-11-11.

The public peer. The public site_b is the dataset edition v2026.11.01 of Artist-Alley-Org/artist-alley-dataset-site-b. It was published on 2026-11-13 as a public prerelease; its tag resolves to metadata commit 9fa8fb76cf208d2b3e6a40d13ed6a44a715ef36e.

  • Contents: 503 assets and 200 posts. 500 images come from five Kenney packs, each with its CC0 notice. The other 3 are the declared overlap rows.
  • Build: generated by seed/scripts/build_public_peer_candidate.py and shipped as a versioned GitHub Release (ADR 0100).
  • Overlap with site_a: the two overlap anchors are byte-identical to their site_a counterparts under distinct asset ids, and the non-overlap control is present only in site_b. These fixtures are declared in seed/contracts/peer-overlap.v1.json and were verified on the published downloads.
  • Where it is used: the pins key site_b in seed/contracts/ci-datasets.v1.json. That pin feeds the nightly’s studio-b peer and the full-corpus scale check.

The former private corpus stays private. The Layer B records that the 2026-07-12 decision describes are not part of the public peer.

  • They are in no public repository, release asset, CI artifact or log.
  • No workflow reads them: the SEED_SITE_B_PATH mount was removed when the UI workflows moved to the public editions (ADR 0100).
  • The Layer B rule, “never published”, still governs them.
  • No CI pin names them site_b.

How the gates were met. The proposed text made the cutover conditional on a per-asset rights audit, the fixture contract, byte verification and the release procedure.

  • The rights audit was not completed in that form. On 2026-11-13 the owner reviewed the recorded source evidence and accepted it for edition v2026.11.01 as a provisional public prerelease. See ADR 0100, “Owner publication decision”. This is not legal rights clearance:
    • all 503 per-file evidence rows remain pending_review;
    • the comparison against Kenney’s current first-party packs was not run;
    • these caveats stay disclosed in the release notes and in the dataset’s README.md.
  • The fixture contract and byte verification were completed. peer-overlap.v1.json holds the overlap fixtures, and the published release was verified end to end against the pinned release.json SHA-256.
  • The release procedure was completed, through ADR 0100’s gates.

The tier rule as amended.

  • Public CI and federation testing may use site_a and the public site_b edition, each under its own recorded per-source terms.
  • Marketing screenshots, the README and the hosted demo keep the Layer A rule above.
  • A later public site_b edition is a new version and goes through ADR 0100’s gates again.
  • The site_a rules, the source-specific attribution obligation and the Kaggle demo-seed steps in RELEASING.md are unchanged.