v1.7.0 — the entity wave, the variant layers, and the replay engine
20 October 2026 · Nabu news
Version 1.7.0 gathers a month of waves — and was cut the day the whole library finished replaying through its new engine: 2,854,118 documents / 128,499,000 passages across 327 live sources, every number on this site regenerated from that fresh catalog (as of 7 November 2026).
The library now answers who. A person index joins eight authority traditions under one roof — the China Biographical Database beside the Epigraphic Database Heidelberg’s Roman names (93,643 persons), the Elephantine papyri’s Aramaic archive (7,577), Japan’s Aozora and Kokubunken rosters, and Korea’s dynastic records — each person carrying its authority’s own identifier, never a guessed merge. Beside it, the facet layer grew two new grains: passages and dictionary entries now carry structured facets exactly as documents long have, so a Sanskrit verse’s meter or a lexicon entry’s register is as queryable as a document’s genre.
The largest dating fill in the library’s history. The Open Korean Historical Corpus — over a million documents of munjip collected works — went from 41% to 95% dated (1,139,409 of 1,198,779 documents, 20 October 2026) by joining each collection to its author’s life years through the Institute for the Translation of Korean Classics’ own bibliographic catalog. Around it, a season of ruled tables: script-tradition bands date the Old Arabian corpus (36,651 inscriptions), Vedic text strata date the Digital Corpus of Sanskrit (15,883), archive and regnal tables drain the Achaemenid shelf, sixty-seven sourced ruler rows extend the cuneiform regnal table (each accession year citing its published chronology), CDLI’s “year after” marks are now read as the year-name tables themselves number them, and a 248-era nengō table converts Japanese era imprints to calendar years at parse time. Korean historical language stages (Middle and Early Modern Korean) are live on the lect axis.
Early-modern Europe is on the map. The imprint lines of 60,325 EEBO-TCP prints — 1,676 spelling variants of what is overwhelmingly a few dozen towns — now read to their bare town names, each matched to GeoNames, Getty TGN and Wikidata identities in the places registry: 93% of the corpus carries a place reference (20 October 2026), with uncertain imprints kept honestly uncertain. The cuneiform world grew the same way: 272,588 CDLI provenience references through the registry’s new decision rows.
The variant readings arrived. Two editorial apparatus layers that
upstream encodes but no reader saw: the documentary papyri’s
<app>/<rdg> variants (about 48,500 entries across 13,700 documents,
editor-attributed) and the Chinese Buddhist canon’s stand-off
witness apparatus — 924,766 entries resolved onto their exact
passages with zero unresolved anchors and the reading text proven
byte-identical. The Pāli canon joined with a Pali Text Society page
concordance whose keys match the held CBETA translation’s own
anchors on 96.5% of pages — two independent encodings of the same
scholarly convention, agreeing in the catalog.
And the engine underneath was rebuilt. The loader now writes multi-row batches (four to ten times faster at library scale), the link-mining layers stamp their work so a four-hour mining pass is paid once and carried thereafter, parsing fans out across worker processes with byte-identical output, and the schema shed several gigabytes of redundant indexes. A guard pack rides beside it: run records that cannot claim success before the work is done, status surfaces tested against freshly born databases, an encoding belt at the adapter read seam, and commit hooks that keep conflict markers and workflow residue out of the public tree. This release’s numbers are the engine’s first full harvest: the entire library — every source, every axis, every index — regenerated from the permanent folders in one pass.
As always: extracted, never guessed; uncertainty is a recorded state, not a missing value; and every figure above carries its date.