Rights

Rights are RebbeHub's biggest risk, so every file carries a rights state and every export passes a gate. Code: packages/model/src/rights.ts, packages/core/src/files.ts, packages/mirror/src/gate.ts.

States#

Where a file starts#

The stricter of what its licence allows and what was decided for its source - the same answers as Sichos-Kodesh's per-edition gate, whose four decisions map one to one (ship→open, ship-with-credit→credit, link-only→link, local-only→preserved):

• public domain, CC0, facts-and-links → open; CC BY and CC BY-NC (Sefaria) → credit;
• free-to-read, site terms, unknown → link; commercial → preserved;
• HebrewBooks → link whatever else is said (a copy of a scan the jobs draw a cover from is kept in the preservation bucket, never served);
• chabadlibrary.org's texts → credit: each page's words are kept and shown, credited to the library, with a link to its page there (a steward's decision, 2026-09-28);
• the Igros app's files → preserved (Sichos-Kodesh decides its own apps);
• hanachos and publisher scans → link (a copy preserved);
• the old typewritten Sichos Kodesh hanachos (5710-5741) → open: the chozrim wrote them under no organisation, the typewritten set was printed privately in 1985, and nobody holds rights in them (the re-typed edition published since 1998 is a publisher scan);
• teshuros, usually printed for free distribution → credit, with a fast path for families to ask for a takedown;
• anything in a locked set → preserved.

Uploads#

"Add a recording" (farbrengen pages) and "Add a scan" (sefer pages) take a file with a rights statement, which sets its licence:

The bytes go to the public bucket only when the state may be served, and otherwise to the preservation bucket, which the API writes and never reads (services/api/src/uploads.ts). A file already known is not taken twice: the uploader is shown where it is. The file joins the catalog through a suggestion, reviewed like any other.

/add (and "Add a hanacha" on farbrengen and sicha pages) takes a new hanacha, recording, or sefer, letter or document the same way. A hanacha's PDF whose uploader is not sure, or that was printed for free distribution, is kept privately (link) until a steward decides; one given freely or in the public domain is served and linked from its farbrengen's page. A hanacha's words follow the same statement: given freely or public domain, shown; otherwise kept and withheld.

A sefer's cover is drawn from the title page of one of its PDFs, as a derivation of it: a PDF the site serves first, else one it only links to (the Otzros library on Drive, HebrewBooks). We store everything and link in public (the owner's decision, 2026-09-29): a linked PDF is fetched from its link and kept in the preservation bucket, never served; its cover is our own picture and is served, credited to the PDF's source where it has a credit, and the sefer's page links to the source for the PDF itself. A PDF in a state that keeps no copy is not fetched (the job says so). The cover is shown only while its PDF is served or linked: a takedown of the PDF (preserved) takes the cover down, and clearing it again brings the cover back.

A new teshura ("Add a teshura" on the Teshuros set, or a scan the upload check takes for one) is always a teshura scan: credit, credited to the families ("משפחות כהן – לוי"), whatever rights statement comes with it. Page images and thumbnails made from a scan are derivations and follow its state.

A family's request#

Every teshura page has A family's request: no account, works without JavaScript, captcha and rate limit as for reports. The teshura's served scans (and their page images) move to preserved at once, as the system actor in the audit log; a rights report and a family_request row are kept for the stewards, who may restore the state with setRights if the request was not from the family. Nothing is deleted (familyRequest in packages/core/src/print.ts). Requests about anything other than a teshura use the general report form.

Translations#

"Add a translation" on a unit's page takes the words themselves, so it takes only what may be copied, and says whose it is:

A publisher's all-rights-reserved translation (Kehot's, a site's terms) is never pasted in: it is listed as a copy elsewhere, a link. A machine translation says which tool made it and is marked as machine text, paragraph by paragraph, until a person checks each one.

Changing it#

Only stewards change a file's state (setRights), and every change is in the audit log. A takedown moves a file to preserved: it stops being served at once and is never deleted. Anyone asks for one at /takedown, with no account; a steward answers within two weeks and takes each file the request points at down in one click from /admin (accounts).

Exports#

The git mirror and the dumps carry catalog facts always, file hashes but never files, and words only when their rights allow: a text copied from a source keeps that source's licence (site-terms and commercial texts are listed as withheld), community text is CC BY-SA, and OCR pages follow their scan's file. The Parquet dump carries exactly what the SQLite and JSON Lines dumps do.

Machines that read the words (search by meaning, citations) send words to Workers AI only when they may be exported; citations found are facts (a link and the reference as written), so they are proposed from any text.