# Rights-safe media library A small, git-tracked corpus of **clearly-licensed** test images with full provenance. It exists so demo pages can ship real sample inputs without hotlinking someone else's server and without bundling anything whose license we can't point at. The initial gallery is curated for the **2D→3D / removable-object family** — depth estimation, matting, semantic/instance/panoptic segmentation, SAM (point/box prompting) and CLIPSeg (text-prompted segmentation) — where the existing per-model pages only had three samples each. The set is deliberately varied: a human portrait, animals, a deep landscape, a busy street, a market crowd, transparent glass, a backlit silhouette, a flat texture, a chart, a night scene, thin structures (a bicycle), and web-page screenshots — including the easy cases and the known-hard ones. ## What's here ``` media/ manifest.json — the source of truth: one record per asset, with full provenance manifest.schema.json — JSON Schema for a manifest asset record assets/.avif — optimized, modern format (smallest) assets/.webp — optimized fallback assets/.jpg — universal fallback ``` Every asset is stored **locally** (no hotlinking) and shipped in three formats: **AVIF** (best compression), **WebP** (broad support), and a **JPEG** fallback. Sources are rendered/optimized to a max of ~1600px on the long edge, sRGB, metadata stripped. ## Provenance rules (non-negotiable) Each `manifest.json` asset records **source URL, creator, license (SPDX or explicit), retrieval date, local path(s), dimensions, and a one-line description** — plus a ready-to-render `attribution` string. 1. **Only clearly-licensed assets are bundled.** Accepted: CC0 / Public Domain, CC-BY, CC-BY-SA (Wikimedia Commons); the Unsplash License (Unsplash); and Paul Kinlan's own sites (owner permission). If a license is unclear or dubious, **skip the asset and record the gap** in `manifest.skipped` — never bundle on a guess. (Example: a Commons file tagged "public domain" whose author died recently was dropped as a likely mistag.) 2. **Store locally.** Demos reference `media/assets/…`; they must never hotlink the original source. 3. **Attribute when required.** CC-BY / CC-BY-SA assets must be shown with the recorded `attribution` when displayed. CC0 / Public Domain need none, but keep the record anyway. 4. **Keep provenance with the bytes.** If you add an asset, add its full record. If you can't fill in the source/creator/license, it doesn't go in. ## Sources used - **Wikimedia Commons** — CC0 / Public Domain / CC-BY / CC-BY-SA. License, author and description are pulled from the file's `extmetadata` via the MediaWiki API, so the provenance matches the file page. - **Headless-Chrome screenshots** — Paul Kinlan's own sites (`paul.kinlan.me`, owner permission) and openly-licensed article pages (Wikipedia article body is CC-BY-SA 4.0; trademarks/logos remain their owners' marks). Captured with `google-chrome-stable --headless=new`. - **Unsplash** — Unsplash License (none included in this first wave; the manifest schema supports it). ## How demos should reference the library (future retrofit wave) This wave only **builds** the library — it is intentionally **not** wired into any of the current demo pages yet (that's a later retrofit). When a demo does consume it, the pattern is: - Read `media/manifest.json`, filter by `familyRelevance` (or `id`), and build a `` so the browser negotiates the best format it supports — AVIF → WebP → JPEG: ```html Busy street scene — panoptic segmentation / monocular depth test input ``` Always set `width`/`height` (from the manifest) to reserve space and avoid layout shift, and render the asset's `attribution` next to it whenever the license is CC-BY / CC-BY-SA. ## Regenerating / extending The curation helpers live in `scripts/` and are one-off (not gates): - `node scripts/fetch-commons.mjs` — search Commons, verify license via the API, pull web-sized renderings of licensed originals. - `node scripts/fetch-screenshots.mjs` — capture web-page screenshots via headless Chrome. - `node scripts/build-media-manifest.mjs` — optimize every source to avif/webp/jpg and emit `manifest.json` with provenance. To add an asset by hand: drop the optimized `avif`/`webp`/`jpg` into `assets/`, then add a complete record (all provenance fields) to `manifest.json`. Incomplete provenance = don't add it. ## Site-wide image ledger (every file, not just this library) `manifest.json` is the **source-asset** library. A separate, fail-closed ledger, [`../image-provenance/ledger.json`](../image-provenance/ledger.json), inventories **every raster image file served by the site** by content hash — the documented sources here, the cropped/re-encoded copies demos use, first-party generated samples (illustrations, mock documents, synthetic textures), and QA screenshots — and maps each to its provenance. The gate `node scripts/check-image-provenance.mjs` (in `deno task gate` + CI) re-hashes every image and **fails** on any file that isn't ledgered or any identifiable person that isn't rights-cleared, so a new image cannot ship without recorded provenance. Faces and people are always clearly-licensed (Commons PD/CC0/CC-BY/CC-BY-SA) or documented assets here — never of unverified origin. The full record renders at `/image-credits/`; CC-BY/CC-BY-SA images shown on a page get a visible credit via `public/image-credit.js`.