Corpus-425

Corpus-425 is a companion test dataset for the World Lattice. It contains 425 books: 25 books from each of 17 literary traditions, selected to see how the Lattice handles books from across the world rather than only from the library it was first built beside.

The outward-facing Python package for using the Lattice is Fihrist. Fihrist ships the artifact, builds the prompts, and validates model placements.

Why the World Lattice exists

The World Lattice was built in response to a recurring problem in book classification: many systems give rich structure to Western or institutionally central material while compressing other traditions into broad regional or identity buckets. The problem is not limited to literature. It appears across religion, philosophy, science, medicine, business, history, and other parts of a library.

Dewey Decimal gives Christianity a large range of classes while placing all Indic religions under a much narrower span. Similar patterns appear when language models are asked to organise books. In the project that led to the Lattice, books by an Indian doctor about genetics and cancer were pulled toward "India" instead of medicine and science. A Japanese book about lean manufacturing was pulled toward "Japan" while operations was missed. The error was not that place never matters; it was that place or identity could overpower what the book was actually about.

The Lattice addresses this by giving a book many possible homes. It has ten independent axes: Discipline, Form, Genre, Place, Time, Language, Agent/Entity, Action/Process, Thing/Matter, and Ethnic/National-group. A book can therefore be a novel, a work of Japanese literature, a book set in Japan, and a book about libraries or well-being at the same time. The goal is not to force a book onto one shelf, but to describe the different ways it belongs.

Why this corpus exists

Corpus-425 was built as a test dataset for that frame. If the World Lattice is meant to work across the world, the practical question is simple: what happens when we place books from across the world onto it?

The corpus covers 17 traditions:

Japan, Korea, China/Sinophone, Southeast Asia, Pacific/Oceania, South Asia, West Asia, West Africa, East and Southern Africa, Russia/Slavic, Nordic, France, Spain/Iberia, UK/Ireland, Caribbean, North America, and Latin America.

Each tradition contributes 25 books. The selection is purposive, not random: the goal is to cover range across era, form, sub-tradition, and country, with one title per author. This is not a canon of world literature and not a sample of typical publishing. It is a test set: small enough to inspect, broad enough to stress the Lattice.

Why literature

The corpus is built around literature because literature is one of the hardest subjects to classify well. Fiction often does not behave like nonfiction: its signal may live in form, genre, tradition, language, setting, prizes, author lineages, or the reader's relationship to the work, rather than in a simple subject label. Earlier anti-library work found fiction becoming a flat blob under generic labels, while nonfiction more often produced discipline-level structure.

That makes literature a strong stress test for the Lattice. Forms such as the novel, epic, lyric, or maqama should not belong to one civilisation by default. A modern Indian novel, a Korean novel, or a Latin American novel should be able to attach to its own literary tradition while also sharing global forms and genres with other books. Corpus-425 gives a spread of works across traditions, eras, and forms so that this behavior can be inspected.

What is included

425 books from 17 traditions were placed onto the World Lattice by three different language models. The release contains 10,634 validated placements.

The dataset includes:

• book metadata and set membership
• model placements onto the World Lattice
• validation records showing which Lattice leaves were attached
• zero-attachment records for cases where a run produced no valid placement

Every leaf in placements.jsonl is a real leaf in the Lattice at the pinned hash:

sha256:92f2002270cecdc79bde3ed951de4efce3d67a302904358b32dd0f143c8b1ec9

Invented leaves were rejected at validation and never entered the table; wrong parents were repaired against the polyhierarchy and the repair counted in runs.json.

Notes on the evaluation records

The corpus also includes records from the model runs used to inspect the Lattice. Those records are useful for model comparison and taxonomy-placement evaluation, but the dataset's main purpose is broader: to provide a world-spanning test corpus for the Lattice.

One known contamination is preserved rather than silently removed: claude-sonnet-5 on 09-east-southern-africa has 19 of 25 books with zero attachments. This was traced to a batch that returned valid JSON after one book and stopped. It is a harness failure, not a model judgement. These 19 rows are listed in zero_leaf_books.jsonl.

Files

File	Rows	What
books.jsonl	425	one row per book: id, set, title, author, category
placements.jsonl	10,634	one row per attachment: book, engine, axis, parent, leaf, confidence
zero_leaf_books.jsonl	19	every (book, engine) pair that drew nothing
sets.json	17	the sets, with the two controls flagged
runs.json	3	per-engine run metadata and validation counters

A placement row:

{
  "book_id": "01-japan/b01",
  "set": "01-japan",
  "engine": "opus-5",
  "axis": "Discipline",
  "parent": "East Asian literature (Chinese, Japanese, Korean)",
  "leaf": "Modern Japanese literature",
  "confidence": "core",
  "culture": "Modern Japan (Meiji-Showa)",
  "status": "valid",
  "rejected": false
}

Blindness protocol

Books were scouted by briefs written to be engine-agnostic: the words lattice, classification, taxonomy, parity, axis and leaf appear nowhere in any brief that reached a scout. An instrument that knows what is being measured finds it.

Enrichment used one frozen instruction, byte-identical across all 425 books, with a single substituted assignment block.

Suggested uses

• World-spanning test set for the World Lattice — 425 books across 17 traditions.
• Three-way model comparison on the same taxonomy, books, and prompt.
• Taxonomy-placement evaluation — 10,634 leaf assignments with definitions available in the companion dataset.
• Literature stress test — fiction and translated literature expose problems that simple subject classification often misses.

Limitations

• The corpus is built around literature; it does not test every branch of the Lattice equally.
• One run per engine. No temperature sweep, no repeats, no variance estimate.
• 25 books per set means set-level findings are indicative; the 425-book aggregate is where the dataset carries more weight.
• The corpus is ~70% fiction by construction.
• P2 adjudication records are not included in this release.

Citation

@misc{corpus425_2026,
  title  = {Corpus-425: a world-spanning test dataset for the World Lattice},
  author = {Khandelwal, Bharat},
  year   = {2026},
  note   = {World Lattice sha256:92f2002270cecdc79bde3ed951de4efce3d67a302904358b32dd0f143c8b1ec9},
  url    = {https://huggingface.co/datasets/bk2022/corpus-425}
}

Licensed CC-BY-4.0.