The release also ships the machinery a living library needs when an upstream cleans house: when PerseusDL deliberately retired ninety-six superseded Cicero editions, the library’s withdrawal alarm said so — loudly, correctly, and forever. A reviewed upstream curation event can now be accepted: the alarm quiets to a dated note and re-arms the moment shedding grows past the accepted level. The health board is fully green for the first time in a week, with nothing swept under a rug to get there.

And the derived layers are going public. The nabu-data sister repository is receiving its largest wave since it opened: the complete kind-classifications table (the fourth axis, per-document, with every upstream label in-band), the cuneiform sense glosses sidecar (Wiktionary’s Sumerian, Akkadian and Hittite lanes, under their own share-alike license beside the CC-BY sign table), and re-derivations of the standing catalog-wide datasets — lect assignments, document dates, places — from a catalog two million documents richer than their last cut. Each dataset carries its full derivation provenance, as always: the exact producing code version, the input identities, and a recipe that makes the build repeatable end to end.

The same pass drained the curation worklist. A full audit of every unmapped upstream label folded roughly a hundred more values — Hittite catalog ranges resolved against the corpus’s own sub-corpus tags, the papyri’s German documentary tail, stray prayers, curse tablets and writing exercises that already had homes — and introduced an honest new category: values declared not-genre (a physical-layout tag, a bare copy marker), reviewed and folded to nothing, with the reason on record. The worklist now opens by saying what it is and what to do with a line, and what remains in it is a quarter of what was there before, all of it genuinely undecided.

And the site learned the lesson a reader taught it: the three axis pages — dates, places, kinds — told one story in three places. They are now one Layers page: the shared doctrine first (every layer extracted, never guessed; honesty buckets counted, never hidden; everything derived and rebuildable), then when, where and what kind as sections that compose into one query:

nabu search lugal --place cigs:GIR --from -2200 --to -2000 --kind administrative

Old links redirect; nothing is lost but the repetition.

nabu search "dis manibus" --kind funerary --lang la

The library’s fourth document axis — kind, beside dates, places, and lects — folds each source’s own genre vocabulary onto a ruled cross-corpus class list: 21 plain families (funerary, dedicatory, administrative, legal, letter, lexical, scripture, divination, poetry, historiography, …), each carrying pointers to the international vocabularies that recognize it (the EAGLE epigraphic types, Library of Congress genre/form terms, Getty AAT). No external standard spans cuneiform tablets, papyri, stone, scripture, and novels at once — the survey behind this axis verified that — so the list is the library’s own, and every upstream label is preserved verbatim beside the fold.

Classification is multi-label. An ode to a ruler is poetry AND royal; a cuneiform “Administrative Letter” is both of its words; a verse epitaph carries funerary/epitaph and poetry together. And it is two-grain: --kind divination finds the whole family, --kind divination/extispicy narrows to the liver omens.

The numbers, from the live census (19 September 2026): 1,953,835 documents classified — 85% of the library — drawn from upstream catalogue genres (cuneiform’s CDLI and Oracc, the Heidelberg and Rome epigraphic databases), the papyri’s HGV text types (Quittung, Vertrag, Mumienetikett), the sibu 四部 classes of the Chinese canon, Japan’s NDC decimal codes, Sefaria’s category tree, and whole-collection declarations where a source IS one thing — the Korean dynastic records, the Buddhist and biblical canons. What no rule covers yet sits in an honest, counted unmapped bucket; sources carrying nothing genre-shaped say so through their postures. The full board lives on the new Kinds page, and nabu kind census prints it in under a second.

This phase gave the library something different: instruments — reference machinery that mints no texts of its own, but makes the texts already held answerable in new ways.

The library now knows where China is. The fourth gazetteer joins Pleiades, Trismegistos and the cuneiform site index: CHGIS/TGAZ (Harvard–Fudan, CC0), 81,292 historical Chinese placenames from 221 BCE to 1911, each carrying its hanzi name, pinyin transcription, valid years and coordinates. nabu place chgis:hvd_167661 answers instantly — 大川, a village-town, coordinates and all.

And the first text-mining lane ran over the Chinese canon. With a gazetteer in hand, the library scanned all 4.57 million passages of the Kanripo corpus — texts that carry no geographic metadata at all — for exact Han-character matches against those placenames: 3,771,151 candidate attestations across 10,055 distinct names, each recorded as evidence, never as fact. Precision rules do real work here (single characters never match, over-common words are stop-listed by measurement), and the candidates wait for human review before any document is actually marked as speaking of a place. The Yijing matching 大川 — “the great river” of the hexagram formulas — is exactly why the review step exists.

Persons arrive as an instrument too. The China Biographical Database (Harvard / Academia Sinica / Peking University; CC BY-NC-SA 4.0, confirmed at first sync) — some 658,000 persons of Chinese history, 7th–19th century — is now held and checksum-verified: names, dates, offices, kinship, place associations. What surface it grows into (a person card? person references on documents?) is a decision the library will take deliberately, the way the places program grew from gazetteers.

The Arabic shelf learned its own intertextuality. From the KITAB project’s text-reuse statistics (CC BY-NC-SA 4.0), the library minted 931,943 reuse edges between pairs of held OpenITI works — which books quote, excerpt and rework which, with aligned-passage counts and chronology flags on every edge. Asking nabu links on a held Arabic work now answers with its textual relatives across eleven centuries.

And one small, satisfying surface: the dictionary card now says when a word was first written down. nabu define bába --lang sl ends the Pleteršnik entry with “first attested: Primož Trubar, Katekizem, 1550” — the crosswalk between a 19th-century dictionary and the 16th-century corpus it describes, rendered where you look words up. Alongside it, four more sources joined the dated timeline, and a batch of era-bound performance assumptions was re-measured against the hundred-million-passage reality.

As always: everything runs on one machine, every number above is measured from the live catalog, and the licenses ride each record — the instruments’ non-commercial grants are enforced by the tooling, not by promise.

Classical Chinese, Japanese and Korean now search exactly. These shelves — the Buddhist canons, the Chosŏn state records, the Japanese library, some eighteen million passages — write without word boundaries, which quietly degraded them to second-class citizens of a word-based index. The fix is a dedicated character-pair index: 23,372,988 pair rows over fifteen sources, so a Han or kana query matches exact character sequences with no guessed segmentation anywhere. The six dragons of the Yijing’s first hexagram, asked as nabu search 六龍, answer in two-thirds of a second from the Book of Han’s music treatise and two Buddhist commentaries quoting 乘六龍以御天 — “riding the six dragons to drive across the sky” — three corpora, one query, mid-rebuild.

The full-text engine went on a diet. The index had been storing a private shadow copy of every passage it indexed — an artifact of the engine’s default design, 39 gigabytes of duplication the catalog already held. Rebuilt contentless and compacted, the full-text store fell from 87 GB to 36 GB with search behavior byte-for-byte unchanged. Fifty-one gigabytes returned to the shelf, nothing lost but redundancy.

The library grew an attic — and it’s searchable. This collection never hard-deletes: when an upstream source withdraws a document or a revision prunes a passage, the text is withdrawn, not erased. Now search --withdrawn searches exactly that shadow collection, and every hit says why it left the open shelves — upstream gone, or revision-pruned — so a vanished reading can be found, cited, and traced years later.

And the meaning-search machinery is built. The phase’s largest single piece is the semantic lane: nabu embed distills each passage of the literary core (4.9 million passages across seventeen Greek, Latin, biblical and treebank sources) into a numerical fingerprint of what it says, and search --similar — on the command line and through the MCP server — answers “where else does the library say this?” with ranked, banded, honestly-labelled neighbors: cross-edition witnesses, drifting quotations, paraphrases. The whole pipeline runs on this machine; no text leaves the box. The first store build is a scheduled overnight run — and the lane will get the announcement it deserves, with live walk-throughs, once its vectors actually exist. Nothing on this site is pasted from imagination.

The plain-language write-ups landed with the code: embed.md on the semantic lane, and lemma-enrichment.md on the silver-lemma campaigns whose tool stack was also folded into one-command installs this phase. Meanwhile the Tools and Examples pages have been reworked around the new capabilities — the Tools page now carries the MCP server as its own section, and the Examples page gained the sinologist and the Old English scholar among its walk-throughs.

The epigraphic core is the DHARMA project’s corpora (EFEO/ERC, CC BY and BY-SA): the Corpus des inscriptions khmères — the Cœdès K-numbers, 1,219 editions of pre-Angkorian and Angkorian epigraphy; the Campā corpus, whose Đông Yên Châu inscription is the oldest attested Austronesian text in existence; the Nusantara charters (Old Malay from Kedukan Bukit to the Laguna copperplate, Old Sundanese, and the sīma charters of Java); and the Corpus of Pyu Inscriptions. Beside them stand the Old Burmese stone inscriptions of Bagan (1,121 faces, Myanmar-script Unicode with a full transliteration lane riding every line) and the desk’s literary wing: the kakawin critical editions — Deśavarṇana, Sutasoma, the Pararaton — with each canto’s meter riding its verses, and the Old Javanese Wordnet as glossary.

The find of the phase hid in the Pyu corpus. Old Mon — the language of the Dvaravati world and the fourth leg of the Myazedi quadrilingual — had no digital corpus anywhere; the surveys recorded the gap as absolute. It turns out two Old Mon inscriptions ride the Corpus of Pyu Inscriptions: seventy-nine lines, the language’s first machine-readable text, now sitting in the same library as everything it touched.

The ingest was an argument for defensive parsing. These are working scholarly repositories, and the first syncs surfaced their reality: inscriptions whose line numbering restarts on every face of the stone, a witness’s page-turns threaded through critical apparatus, editors resuming an interrupted line, duplicate stanza numbers in flagship editions, and one template placeholder reading “languageb”. Every shape was either given its honest citation form or quarantined loudly and recorded — nothing silently dropped, nothing silently merged.

The library now holds 2,268,754 documents — 106.1 million passages — in 167 language codes across 25 desks.

The licensing is the quiet achievement. The corpus is one non-commercial dataset, but its records are not one thing: each carries its own copyright status, and the adapter reads it per record — 1,198,691 relabel Public Domain, 15 carry Korea’s KOGL attribution license, and 73 no-derivatives records earn no upgrade at all and keep the stricter class. A million-record source where every card tells the truth about its own terms.

The ingest itself argued for defensive design. The first sync quarantined 2,761 records rather than guess: 2,760 turned out to be English and French documents in the Copyright Commission’s collection — classified and recovered the same day — and one had no body at all. And a silent schema drift (the deposit spells its copyright field differently than the published sample) was caught not by any test but by spot-checking stored cards against the source, then pinned in fixtures with real deposit bytes.

The same phase re-grained three Menota manuscripts whose pages carried unnumbered lines — Old Norse prayer books and legal manuscripts now cited by folio and running line, thirty thousand lines where four hundred page-blocks stood. The library now holds 2,265,769 documents — 106.0 million passages — in 155 language codes. A month ago it had not yet crossed one million.

Each line arrives twice over: the edition’s unified Latin transliteration is the searchable text, while the original script — native Italic written right-to-left, Latin capitals, or Greek letters in the far south — rides every line as an annotation naming its alphabet. The registry minted an Umbrian anchor the same day, completing the Sabellic pair; the comparative desk’s Classical-Latin equivalence keys (from CEIPoM) already reach into the Tables, so search --lemma precor finds pesnimu — labeled as scholarly equivalence, never counted as attestation.

Two quieter milestones close the week. The 2026-08-31 census reads 102.4 million passages across 1,066,990 documents — the hundred-million line crossed without ceremony somewhere between the early-modern presses and the rhyme books. And the rebuild machinery completed its economics arc: derivation stamps, a builder-scoped digest, an owner-attested trust bridge, and an LFS-aware identity for the one shelf whose materialized payloads had read as dirt forever — so the library that once took sixteen hours to rebuild now re-verifies an unchanged catalog in minutes, honestly.

Numbers as of 2026-08-31.

The ingestion respected the edition’s variety the hard way. A first crawl parsed only the corner of the corpus the samples had shown; the full census then taught the parser four page shapes the site actually serves — tabular scores with witness sigla, labeled paragraphs, free prose — and the steady state reads every text page the archive publishes, with three broken-markup pages quarantined by name rather than served with holes.

On the dictionary side, the Coptic lexicon gained its second etymological witness: the KELLIA project’s Egyptian-etymologies table now rides beside the ORAEC crosswalk — two independent scholarly witnesses to each Coptic word’s descent through Demotic and hieroglyphic Egyptian, merged where they agree (they never disagreed: a zero-conflict census preceded the merge) and credited separately where each speaks alone.

Numbers as of 2026-08-31.

At its head stand the Guangyun (1008, 25,336 entries) and Wang Renxu’s Qieyun manuscript (17,227 entries) — the Middle Chinese standard the whole discipline reconstructs from — joined by Fujita Takumi’s restoration of the Qieyun itself (11,158 entries), the lost 601 CE parent of them all, recovered from the surviving fragments and quotations. Then the record jumps six centuries: the 蒙古字韻 (1308, 9,446 entries), the ‘Phags-pa-script rhyme book that writes Old Mandarin in the Mongol empire’s universal alphabet, and the 中原音韻 of 1324 (5,877 entries), the arch-source of Old Mandarin phonology, compiled for opera singers who needed to know what actually rhymed. The registry minted a proper Old Mandarin stage for them — no ISO code exists for it, so the lect layer now carries what the standards cannot.

Beside the rhyme books, three more doors: a Classical↔Modern Chinese parallel corpus — 14,608 chapter documents, 1.94 million passages, ~967k sentence pairs of the classical canon facing its modern rendering, the largest parallel shelf in the library after the biblical hub; the Monlam Tibetan lexicon (449,829 entries), the largest licensed machine-readable Tibetan headword list, which more than doubled the dictionary layer’s total; and DACON, the first annotated corpus of Classical Newar — Nepal’s Sanskrit-cosmopolis literature, gold-tagged — a small shelf (4 texts, 977 passages) that brought a whole language family into the registry.

Numbers as of the 2026-08-31 census; the shelves entered the catalog with the 2026-08-29/30 first syncs and were verified and wired the same week.