作品識別子 / Work identifiers

Every published work has one identifier of the form <work-id>_<card-directory>, consisting of six digits, an underscore, and six digits:

000092_000879

It appears as works[].slug in each signed manifest, as the directory name in the published routes, and as the file name of a withdrawal record:

https://soranoha.org/works/000092_000879/tei
https://soranoha.org/works/000092_000879/markdown
https://soranoha.org/works/000092_000879/plaintext
https://soranoha.org/works/000092_000879/tei-validation
https://soranoha.org/withdrawn/000092_000879.json

Both components come from Aozora Bunko. 000092 is the catalog's 作品ID (work id) and 000879 is the contributor card directory the source archive is filed under, as in cards/000879/files/. Together they name the Aozora Bunko card the text was taken from:

https://www.aozora.gr.jp/cards/000879/card92.html

The card URL drops the leading zeros from the work id; nothing else needs translating.

What the identifier promises

It is stable across releases. The identifier is computed from one source's own coordinates and nothing else, so no change elsewhere in the corpus can move it. Adding works, removing works, reprocessing the corpus, changing the converter, or changing this project's own tooling all leave it exactly as it was.

It denotes the work, not a particular file. If Aozora Bunko repackages the archive, re-proofreads the text, or publishes a corrected edition, the identifier stays the same and the text under it changes. Changes are recorded explicitly rather than hidden. The manifest carries source_content_hash per work over the canonical identity of the source bundle rather than a raw archive hash, so two releases can be compared exactly. Cite the identifier for the work; cite the identifier plus source_content_hash when you need the precise bytes you read.

Withdrawal moves where it resolves. Publication is append-only, but a work can be withdrawn. A withdrawn identifier stops appearing under /works/ and appears under /withdrawn/<identifier>.json, which names the signed governance event carrying the public reason. The releases that included the work keep their manifests, signatures and content; withdrawal removes the work from the current release, and never rewrites a past one.

An identifier would also change if Aozora Bunko refiled the same work under a different contributor card, since the card directory is part of it. Only that work would be affected; every other identifier is computed independently.

What the identifier omits

The identifier omits person IDs, archive filename stems, and format variants (ruby/txt). Those describe acquisition details rather than intellectual work identity. Including them in permanent URLs would couple work identity to edition selection mechanics.

The card directory is not a person ID that happens to look like one. It is the card the archive is filed under, which for 146 works belongs to a contributor other than the first author.

The card directory was kept even though the work id alone is unique across the present corpus: in the catalog this release builds from, no work id is filed under more than one contributor card. It is the documented disambiguator for a work id that is filed under several, so the scheme is collision-safe by construction rather than by observation, which is what a permanent identifier needs.

The identifier and the download filename

Every artifact is served at two locations. Citations use /works/<identifier>/<type>, which can be constructed directly from the work identifier. Browser downloads and curl -O receive /works/<identifier>/<author>-<stem>-<identifier>.<ext>. Both paths resolve to the same blob.

The filename is a convenience, whereas the identifier is the citable reference. The filename is a serving-layer projection. It can be corrected in a later release, its author component selects one contributor out of a work that may record several, and its middle component is Aozora Bunko's own archive stem rather than an identifier assigned by this project. The signed catalog carries the inputs (archive_stem and the romanized name parts) rather than the assembled name, because a derived value stored beside its own inputs can disagree with them and the chain is append-only.

The identifier is a component of the filename rather than an alternative to it. 470 author-and-title pairs are shared by 2357 of the 17810 works in the Aozora Bunko catalog, so a name built from author and title alone would silently collide for more than one work in ten of any bulk extraction. Splitting a filename on its hyphens recovers the identifier as the last component; the bulk archives also carry it in catalog.csv.