The grant harvest and the research bench
13 October 2026 · Nabu news
A hundred thousand documents and a change of character. The last three phases added shelves the way the library always has — but they also added something new: instruments for statistical research across everything the shelves hold.
The whole Tocharian B corpus. The West Tocharian manuscripts sit behind a members-only gate at TITUS, and the library now walks through it politely: a credentials seam stores the granted password once (set interactively, never echoed, never in history) and every fetch carries it. What came back was more than expected. The “Tocharian B” chain turned out to run 4,349 pages — the Berlin Turfan and SHT series, then the London India Office and Paris Pelliot collections, one “Next part” link at a time — and after two rounds of parser work the whole chain parses: 4,250 documents / 43,480 passages, with the data-entry editors of all three collections credited on every card. The honesty ledger is short and documented: eighteen pages stay quarantined, most carrying text with no line citations to mint. And two of them turned out not to be Tocharian at all — Maue’s handlist of Tumshuqese manuscripts claims both leaves (one is the Tumshuqese Araṇemijātaka, and the leaf itself spells the hero’s name), so the library’s first two Tumshuq Saka documents join under their own code.
TAMILEX, a day after the ask. The Hamburg Academy project’s critical editions of the classical Tamil corpus — the Caṅkam anthologies, the Tirukkuṟaḷ with its commentators, Tolkāppiyam, the Nālāyirat Tivviyap Pirapantam — carry a per-word wordsplit layer (segmentation, gloss, grammatical role, clitics) that exists nowhere else for Old Tamil. The maintainers confirmed CC BY-NC-SA for the whole estate, and the next morning the library held 7,830 documents / 14,972 passages with 262,093 gold word-level rows mirroring the project’s own indexer — and re-syncs tracking upstream’s practically daily edits.
New scripts on the shelves. The Balinese Wikisource’s lontar transcriptions arrived manuscript-direct — 1,252 palm-leaf manuscripts in Balinese script, no printed edition in between — beside Gujarati bhakti prints, Jawi-script Malay manuscripts from Leiden and the British Library, and the Thai Collected Chronicles. Then the Gurmukhī canon: the Shabad OS database’s 141,264 scripture lines at line grain — the Gurū Granth Sāhib, the Vārāṅ of Bhāī Gurdās — with the registry growing a Punjabi anchor, a granthic Old Punjabi stage for the Sant Bhāṣā written language, and Gurmukhi joining the script checkers. South India deepened too: the DHARMA conversions of the printed South Indian Inscriptions volumes (589 license-marked inscriptions with their original ASI editors credited) and Project Madurai’s 755 classical Tamil e-texts at 1.2 million lines.
The research bench. The reading instruments were always good;
the statistics instruments barely existed. Now: nabu export
writes research tables (documents with their full axes, the links
graph, the lemma attestation matrix) with a license manifest
beside every file — each export names every source it drew on and
its license class, and one flag cuts a provably redistributable
set. nabu lemma answers attestation profiles from precompiled
rollups — first and last century with the date’s provenance class,
gold and silver curves shown separately — and lang --curve and
lang --mixing serve the demography and multilingualism questions
the same way, instantly. A dating lane joined for model output:
estimates live in a shelf behind an owner-controlled gate, project
as visibly-flagged silver dates that never mix into surfaces
promising upstream dates, and search --date-basis filters by
provenance class — the holdout discipline for dating research as a
flag. And the text-reuse machinery grew tuned per-language
profiles with resumable chunking; along the way the census found
that Chinese parallels had been silently empty (character n-grams
never matched the index’s unbroken runs — fixed through the CJK
bigram index), and the library’s first full-scale Latin reuse
run — the classical canon probed against the Patrologia — is
mining as this post goes up.
The ledger line: the library stands at 2.37 million documents, 112.5 million passages, 208 live sources across 28 research desks — a quarter-million documents and three writing systems richer than it was ten days ago, with every grant, census and quarantine recorded where a reader can check it.