Known gaps

Honest limits of release 2026-09.2.

• Books changelog is empty. First books load (2026-09.2). GET /v1/changes?dataset=food has 127,333 updates from the 2026-10.2 pack-shot increment.
• Authors are sparse. Some editions still return authors: []. Open Library keys are resolved to names on read when we can.
• Subjects and edition metadata are sparse. publisher, publication_date, page_count, format are often null.
• Synopses cover a slice. About 8.4M of 78.2M works have description.
• Split works. Open Library split some popular titles into several work ids. Match and search collapse those when the title is an exact hit.
• One source. Open Library only. Library of Congress and BookBrainz are not in this release.
• Recommendations are content similarity, not other readers’ shelves. About 1.28M ISBN+synopsis works have embedding neighbors (~64M rows). Open Library ratings/reading-log counts bias ranking; that is not collaborative filtering. Seeds with no neighbor rows return { results: [] } — no subject-keyword fallback.
• Genre browse uses edition subjects. genre=fantasy works when subjects mention fantasy/magic/etc. Fiction, general does not map to a facet.
• Work merge is incremental. Split fds_wk_* collapse in the fetch window today; aliases land in hash buckets, not a 78M re-ER.
• Resale is prohibited. Bulk dumps are for your own systems.
• Food images are a GTIN join. 127,333 of 440,110 branded foods have at least one OFF pack shot (257,715 JPEGs: front / nutrition / ingredients). Unmatched barcodes have no photo. OFF photos are CC BY-SA; USDA nutrition is CC0. We do not ingest the Open Food Facts product database.
• Branded nutrients are label values, not lab assays. FNDDS and SR are per-100 g composites. Some FDC sodium values are implausible; those rows carry quality_flags.
• One food per GTIN. Older label versions stay in the R2 lakehouse only.
• Movies & TV posters prefer Commons P3383, then Wikipedia infobox images (page_image), then Posterflix/TMDB (tmdb_display) when Wikipedia only has a logo. Those last two are display/identify only. Popular film/series posters are hosted at /media/titles/{id}/poster.{400,1080}.jpg. Person portraits remain Commons P18. Wikipedia summaries are CC BY-SA — attribute summary.attribution_url wherever the extract is shown.
• Episode coverage follows Wikidata. Long-tail series are thin; series counts.episodes can exceed the number of episode rows. Animated seasons (Q125354488) and animated episodes (Q116048824) were dropped from the first dump pass. Normalize now fetches missing season rows and season P527 episode QIDs without a dump re-ingest. TVmaze stills fill season/episode art on read when the parent show has a TVmaze id and season_number is set.
• Season pages are first-class. Series pages list seasons when we have fds_se_ rows; click through for that season’s episodes. Flat episode shelves remain only when the series has no season rows.

What is solid

• ISBN-10 / ISBN-13 lookup
• Stable ids + 301 redirects
• Title resolution with article/apostrophe fold and summary/study-guide demotion
• Author search and content-based recommend
• MCP and REST sharing one key
• Check record: 9780593135204 Project Hail Mary