bk2022/world-lattice

World Lattice

A ten-axis faceted classification of 3,660 leaves, built to place any book — science, religion, philosophy, medicine, literature — without treating one civilisation as the default and the rest as local colour.

sha256:92f2002270cecdc79bde3ed951de4efce3d67a302904358b32dd0f143c8b1ec9

Use it through Fihrist, the Python package that ships the Lattice artifact, builds the prompts, and validates model placements.

The problem

Open Dewey Decimal, the world's dominant shelving system, at religion:

	Dewey range
Christianity	200–289
All Indic religions	294 — Hinduism is 294.5
Shinto	299.5 — where Japan's own NDC gives it a full 170 division

Ninety classes against one decimal place. Library of Congress encodes the same thing structurally rather than numerically: the Americas get two top-level classes; Asia and Africa share one.

This is not a historical curiosity. An LLM asked to organise a personal library reproduces it from training data. The project's own records show it from day one — an Indian doctor's genetics books filed under "India," a Japanese lean-manufacturing book filed under "Japan" with Operations lost entirely. Both caught by human correction, not by the machine.

Two distinct failures, and they need different fixes:

	Failure	Cost
Identity magnet	author identity substitutes for content	a wrong home
Completeness	one salient home hides the others	a missing home

The mechanism: the false universal

The deeper finding is not that non-Western material gets fewer slots. It is that the West is encoded as the unmarked universal while every other tradition is marked.

There is a "Chinese scientific tradition." There is no "European scientific tradition" — Greek astronomy and the Scientific Revolution simply became "science." The framing says real knowledge is universal, and other cultures have quaint traditions.

An audit that counts leaves cannot see this. Counts can be perfectly equal while a shared global form sits inside one culture's branch. It is visible only structurally, and the fix is one line: mark the West too.

Five branches carried the pattern — Literature, Science, Medicine, Law, Education. Art, Mathematics, Religion and History already did it correctly and became the in-house template: a universal parent with every culture's contribution, including an explicitly named Western one, as parity leaves beneath it.

The result

Religion — the case Dewey fails worst:

Christianity	Islam	Hinduism	Buddhism	Judaism	African traditional	Afro-diasporic
12	12	11	10	10	8	6

Medicine — note the marked sibling:

Clinical specialties (biomedicine)	Ayurveda	TCM	Unani/Tibb	Western humoral / Hippocratic-Galenic	Kampo	Siddha	Sowa Rigpa
25	12	10	8	7	5	4	3

Philosophy — "Western philosophy" is a named sibling, not the field itself:

Western	Indian (Hindu)	Islamic	Buddhist	Chinese	African	Japanese	Korean
23	12	11	11	10	6	6	4

Mathematics — Greek mathematics as one tradition among several:

Classical Greek · Classical Indian · Classical Chinese · Medieval Islamic · Japanese wasan · Mesoamerican · Ethnomathematics

Painting — the template the rest was rebuilt against:

Oil · Fresco · Watercolour · Chinese ink-wash (shan shui) · Chinese bird-and-flower (huaniao) · Nihonga · Persian & Islamic miniature · Mughal miniature · Indian folk (Madhubani, Pattachitra, Warli) · Korean minhwa · Thangka · Aboriginal Australian dot

Dewey's Hinduism-as-one-number is dissolved, and it is dissolved the same way in five domains at once.

Literature: the hardest case, and the last one fixed

Literature resisted longest, and the fix required leaving library science entirely.

A slot test in July 2026 passed on religion, philosophy, medicine and science — and threw up one defect: Indian novels attaching to Novel under a parent literally named "Western & Greco-Roman literature." Three stacked problems:

1. Form conflated with tradition. The branch listed forms under tradition parents while a separate Form axis also carried Novel. Form was represented twice.
2. The global novel filed as Western property — the classic Watt error. The novel was adopted worldwide; Genji predates Watt's "rise" by seven centuries, and xiaoshuo and maqāma are independent lineages.
3. No modern layer. Non-Western traditions had classical nodes only (Kavya, Itihasa, Ghazal, Waka). A modern Indian novel had nowhere native to go even after fixing #1.

The correction came from literary studies rather than shelving practice — the field had solved this decades earlier:

• Gérard Genette, The Architext (1979) is about precisely this confusion. It separates mode (narrative / dramatic / lyric), genre (novel, tragedy, epic) and form (verse, prose, fixed forms) as orthogonal axes. Frye's Anatomy of Criticism is likewise four independent grids, not one tree.
• Tradition = region × era, at parity. The marquee world-literature theories (Casanova, Moretti) are themselves core/periphery Eurocentric; parity comes instead from practice — the Norton Anthology varying its structure so Europe is not first — and from area studies, which frames Tagore and Lu Xun as the modern layer of their own tradition, not as Western imports.

Both fixes are in the artifact and checkable:

Novel now lives only on the Form axis, under Prose fiction forms, with 29 peers:

Novel · Novella · Novelette · Maqama · Monogatari · Dastan/Qissa · Hikayat · Katha · Chuanqi · Huaben · Biji · Zhiguai · Saga · Picaresque · Epistolary novel · Vernacular novel (baihua xiaoshuo) · Frame narrative · …

No tradition owns it. Poetic forms are split the same way — and one of the parents is Poetic forms — Western.

Each literature tradition now runs region × era, with the modern layer present:

South Asian	European / Western	African
Classical · Medieval and Bhakti · Early-modern · Modern · Contemporary	Classical Greco-Roman · Medieval · Early-modern · Modern · Contemporary	Ancient and classical · Medieval and Islamic-era · Oral and epic · Modern · Contemporary

Ten literature parents, at 4–10 leaves each: East Asian 10 · Arabic/Persian & West-Asian 8 · African 5 · Central Asian 5 · European/Western 5 · Jewish/Hebrew 5 · Latin American 5 · South Asian 5 · Southeast Asian 5 · Indigenous & Oral 4. "Greco-Roman" is now a leaf inside the Western parent rather than the unmarked default.

A book has many homes

The second half of the problem. Ten axes are independent; a book attaches to leaves on as many as its content earns, and to none where none apply. Clusters are computed views over those attachments, never stored categories.

This is what lets a book be two things at once. From the companion corpus, a real placement:

What You Are Looking For Is in the Library — a novel Discipline → Contemporary East Asian literature · Positive psychology & well-being Form → Novel · Genre → Literary Fiction, Inspirational Fiction Place → Japan · Agent/Entity → Library/museum/archive, Older persons

Fiction is not quarantined from knowledge: under the strongest engine measured, 100% of fiction books also carry a Discipline attachment, and most books touch five to eight axes.

Three axes exist specifically to carry connections no discipline tree can make — Agent/Entity, Action/Process, Thing/Matter. Self-cultivation joins Taoism, habit formation and Stoicism, which Discipline files under three branches that never meet. A book on sport reaches human performance the same way.

Axes

Axis	Leaves	Answers
Discipline	1,246	what field of knowledge
Action/Process	467	what happens
Agent/Entity	438	who or what acts
Thing/Matter	364	what physical subject
Genre	245	what kind of book (fiction only)
Form	221	what shape the text takes
Language	216	what language
Ethnic/National-group	181	whose group
Place	175	where
Time	107	when
Total	3,660

Lineage — a parity-audited mosaic

The design borrows; it does not invent. Five parallel surveys covered ten regions and some fifteen national systems, with a sixth pass on France and Italy. The core finding: every culture that classified its own material achieved parity, by one of two routes — faceting, or deep local enumeration. Neither Dewey nor LCC fixes the bias, because no taxonomy is neutral; each centres its author's world. So the leaves are sourced as a mosaic of native structures, unified by a parity audit — the one component that inherits no worldview, because it only measures asymmetry.

Three systems shaped this artifact most:

India — Colon Classification (S. R. Ranganathan, 1933). The origin of faceted classification. PMEST — Personality, Matter, Energy, Space, Time — composes a subject from independent facets rather than forcing one shelf position. Religion, language and place become separate dimensions, which is structurally why facets deliver parity: nothing has to lose to let something else win.

Italy — Nuovo Soggettario (Florence) and the Integrative Levels Classification (Gnoli, ISKO Italia). Nuovo Soggettario is faceted subject indexing built independently of Ranganathan, and its macro-facets are Agenti · Azioni · Cose · Tempo, plus place. Agents, actions, things, time. That is PMEST again, reached separately in Italy — and it is this Lattice's Agent/Entity, Action/Process, Thing/Matter, Time and Place. Two faceted traditions, built a continent and half a century apart, converge on roughly the same five fundamental axes. Italy also supplies the architectural precedent: a faceted subject layer where parity is recovered, decoupled from a Dewey shelf layer that stays biased. ILC's phenomenon-based stance was examined and not adopted — this Lattice stays discipline-based.

Finland — Kaunokki, KAUNO and BookSampo. Fiction indexing exists because fiction resists subject classification, and the Nordic lineage went furthest: Pejtersen's Danish Book House (1970s — subject matter, setting, author's intention, and accessibility), Saarti's Kaunokki thesaurus (1996), the KAUNO ontology, and BookSampo, run by Finnish public libraries as linked open data for roughly two million users a year. Finland also demonstrates the enumeration route: YKL forked Dewey to keep local depth rather than accept its compression.

Two negative results kept the design honest. Sweden switched from SAB to Dewey in 2011 and lost Swedish resolution — research libraries had to bolt on local expansions, public libraries refused to leave SAB. And Latin America adopted faceted CDU but never authored local depth, staying coarse: faceting is necessary, not sufficient. Someone has to do the work.

Parity audits

Nine audit records sit in auditReports. Six carry a real finding; three are stubs with report: "test".

Axis	Verdict	Leaves added
Language	asymmetry found	72
Place	FAIL	39
Action/Process	asymmetry found	38
Agent/Entity	asymmetry found	30
Discipline	asymmetry found	26
Ethnic/National-group	asymmetry found	18
Time · Thing/Matter · Form	not audited (stubs)	1 each
Genre	no record at all	—

Read as coverage, not quality: four of ten axes have no parity evidence. They may be fine. Nobody checked.

Two dimensions are required, and the second was learned the hard way:

• Count parity (build-time) — does each culture get comparable resolution?
• Structural parity (post-hoc) — are shared global forms placed neutrally, and is form kept separate from tradition?

The Lattice once passed a count audit while filing Novel under "Western & Greco-Roman literature." Equal counts, biased placement.

Files and schema

Single file: world-lattice.json, with world-lattice.sha256 alongside. Nested JSON, not tabular — the dataset viewer will not render it usefully. Read it as a document.

{
  "id": "discipline.biblical-studies",
  "name": "Biblical studies",
  "parents": ["Christianity"],
  "definition": "Scholarly study of the Old and New Testament texts, exegesis, and canon.",
  "sourceSystem": "LCC (BS) / DDC 220",
  "culture": "Christian"
}

parents is a list — polyhierarchy, not a tree. definition is present on every leaf and is what makes automated placement checkable: an assignment can be argued against a definition, not merely a label. sourceSystem and culture are optional; roughly 3,000 of 3,660 leaves carry no sourceSystem at all.

Known gaps

• No Pacific / Oceanian literature parent, and no Caribbean literature parent.
• Four axes carry no parity evidence (Time, Thing/Matter, Form, Genre).
• A handful of near-duplicate leaves survive a light dedup (Fable / Fable / Parable; Chuanqi / Chuanqi (Tang tale)).
• Parity audits share blind spots with the models that run them. Psychology was found missing after an earlier pass had declared the Lattice corrected. Agreement is not proof.
• The parity-rich subfield lists are a virtue strong models use well and weak ones abuse as a checklist to complete — an observed failure mode, concentrated in big soft-boundaried fields.
• Closing any gap is an amendment, and amendments run both audits. Not a quick edit.

Lineage note

sourceSystem strings reference LCC, DDC, BISAC and reference works as descriptive provenance — where a term is conventionally found. No classification codes or scheme structures are reproduced. The hierarchy is authored.

Using it

pip install fihrist

fihrist verify

The package ships this artifact, pins the hash as a compiled-in constant, refuses a drifted copy, builds the placement prompt, and validates model output back against the frame — invented leaves dropped, wrong parents repaired.

See also bk2022/corpus-425 — 425 books from 17 traditions placed on this frame by three engines, with the pre-registered probes that judged the result.

Citation

@misc{worldlattice2026,
  title  = {World Lattice: a ten-axis parity-audited classification frame},
  author = {Khandelwal, Bharat},
  year   = {2026},
  note   = {sha256:92f2002270cecdc79bde3ed951de4efce3d67a302904358b32dd0f143c8b1ec9},
  url    = {https://huggingface.co/datasets/bk2022/world-lattice}
}

Cite the hash with the name. The frame goes coarse → fine → stabilize, and a claim about "the World Lattice" without a hash will not stay true.

Licensed CC-BY-4.0.