Doubt made visible, and the vernaculars arrive

18 August 2026 · Nabu news

Two waves since the layers learned to join, one epistemic and one territorial.

The library now says how sure it is — and only when it isn’t. The uncertainty doctrine has three house tiers — certain · probable · uncertain — with two rules that keep it honest: absence is not a tier (an undated document stays honestly undated, it is not “uncertain”), and precision is not certainty (a wide date band is wide, not doubtful). The render axiom does the work: certain is silent, deviation is labeled, and the upstream’s own word always rides verbatim beside the house word. A place card says probable ("presumably") — the gloss and the source’s hedge together, a render-time label that never rewrites what upstream said. The same object rides the MCP payloads, merged only when something is actually below certain, so machine consumers pay nothing for confidence.

The medieval vernaculars arrive. The largest wave of new languages since the library opened:

Old Spanish, via the Hispanic Seminary’s Old Spanish Textual Archive — semi-paleographic transcriptions of medieval manuscripts kept verbatim (the curly-brace transcription convention parsed, not flattened), with a lemmatized sibling lane. The archive’s own works table resolves each codex’s lengua: alongside the Castilian bulk, the Leonese texts carry Old Asturian and the Navarro-Aragonese ones Old Aragonese — two stages minted for the occasion in the nabu-lects registry.
Old Italian — the full Commedia as a dependency treebank, all 122,038 words, while the big Old Italian corpora remain walled.
Old Swedish — the laws, religious prose and rhymed chronicles of Fornsvenska textbanken, under the new sv:old stage.
The Spanish sonnet corpus — 4,530 sonnets across five centuries.
Achaemenid Babylonian — 2,830 legal and administrative texts from the Persian-empire archives (the Murašû firm of Nippur, the Eanna temple of Uruk), lemmatized by the Helsinki team’s pipeline and carried at the honest silver tier: automatic annotation is labeled as such, never passed off as gold.
An Etruscan concordance — Annie Burman’s crosswalk of fourteen citation systems (Trismegistos, CIE, Rix, TLE, CIL…) joins as a links instrument, so an inscription’s whole identity trail answers from one query.

And the library re-proved it can die — this time all the way. The restore drill (back up the three permanent folders, restore to a bare machine, rebuild everything derived, compare) had never run fully green on the live corpus; its first serious runs found real nondeterminism — two canonical inputs claiming one identity, a catalog’s citation markers minting phantom documents. Those became a five-layer determinism contract, the drill grew a space guard and per-stage narration with honest progress, and the proof now closes: 977,241 documents and 68.6 million passages, derived twice, byte-identical. The catalog is a pure function of the sources, the decisions, and the code — demonstrated, not asserted.

Alongside: graded reading reached cuneiform (search --signset "ŠEŠ, AK" finds real tablet passages a learner’s sign inventory can actually read, the sign-list analog of the Han --charset lane), and the first public hieroglyph frequency table — Gardiner-code counts from a gold-annotated corpus, per genre — joined the open datasets.