The library proves it can die — then publishes its layers

11 August 2026 · Nabu news

The last few days rebuilt the library’s foundations, then published what stands on them. Three phases, one arc.

Everything permanent fits in three folders — proven, not assumed

The derivability law is now an enforced architectural invariant, not a habit: everything permanent lives in three plain-file folders — canonical/ (the texts, the universal asset), config/ (the project’s definition), and local/ (this instance: the owner’s rulings, shelves, notes, and the operational ledger, newly elevated to a first-class folder) — and every database is derived from them. Guard tests now classify every table in every database by that law, census every code path that writes a permanent file, and refuse new ones that arrive undeclared. The stronger claim has its own drill: restore ONLY the three folders to a bare directory, re-derive everything, and compare counts against the live library. Backups stopped being a hope.

The character desks learn to teach

A capability survey from Edubba — the scribal-school sibling — came back as a shipped feature wave the same day. The Han character card grew a graded-reading lane: give it the characters a student knows, and it finds real corpus passages readable with at most N unfamiliar characters, cleanest first, naming each passage’s strangers — backed by a 16-million-row rarest-character index, so the answer is instant at any charset size. Cards gained per-document vocabulary profiles with coverage percentages, JSON output for machine consumers, and a didactic panel: where Edubba’s course material covers a sign, the library’s card now says so, chapter and keyword, with uncertain etymologies flagged in front of the gloss. The school teaches the scripts; the library carries the texts; the seam between them is now data.

The layers go public — eight new datasets

nabu-data grew from thirteen datasets to twenty-one, and the new eight are a different kind of thing: not single-language tables but the library’s own compiled layers across the whole catalog, published under the new mul/ (multiple-languages) namespace:

mul/lect-assignments — 482,287 per-document historical-stage assignments (Old vs Neo-Babylonian, Vedic vs Classical Sanskrit…), each with its basis in-band. No other project publishes stage stratification at corpus scale.
mul/place-refs — 969,929 document→place claims over 7,309 places, every historical ref spelling folded to clean gazetteer identifiers, shipped as four shards.
mul/places-lpf — the same places as a Linked Places Format FeatureCollection plus the LP-TSV upload table the World Historical Gazetteer accepts: attested spellings cited to their corpora, coordinates, aggregated date spans, cross-gazetteer links.
mul/document-dates — 703,372 normalized year spans with the verbatim upstream dating string riding every row.
mul/char-postings — the Han character × corpus doc-frequency census behind the graded-reading lane, 38,397 rows.
sux/sign-table — compiled cuneiform sign cards: OSL identity, print-list concordances, CDLI readings, and real attestation counts — how many held documents actually use each sign.
egy/unikemet-signs — the Egyptian hieroglyph spine: 5,067 codepoints with Gardiner-style codes and the JSesh/Hieroglyphica/ IFAO concordances every Egyptological tool joins on.

Every dataset slices by license class row-by-row (non-commercial and ODbL inputs excluded and censused, never silently dropped), and share-alike datasets now carry their own LICENSE file in-directory.

And one upgrade shipped on the way out: cuneiform tablets dated by regnal year (“Šulgi year 44”) now resolve through a ruled king table — middle chronology, stated in-band — so 95,054 CDLI documents moved from ±100–300-year period bands to single-year dating, 82,133 of them to an exact conventional year. Rulers whose absolute chronology is genuinely uncertain stay period-banded on purpose: the table’s header names them, and false precision is refused.

The v1.1.0 release of nabu-data — the first carrying the cantigas — cuts alongside this wave.