The lect layer stabilizes — and becomes part of the core
6 August 2026 · Nabu news
A layer becomes infrastructure the day its defects are treated as infrastructure defects. This phase did exactly that to the lect layer, in five moves.
No dating runs backwards anymore. Last phase’s audit caught one inscription whose dating interval was stored reversed — and the sweep behind it found 82 such rows across six sources: BCE years with dropped signs, bounds in each other’s seats, one upstream digit typo. Each extractor now repairs its own defect class mechanically (signs and seats move; digits never do — upstream typos surface in the raw dating, which always rides), a standing health check reports any future reversed interval by name, and the re-run date inference recovered the affected documents with zero refusals.
The deferred mappings closed. New Kingdom Egyptian is now a stage
of its own — the registry’s band deliberately overlaps Middle and Late
Egyptian, because a corpus dated “NK” straddles that transition and an
honest coarser claim beats a confident wrong one. The Dead Sea Scrolls’
Aramaic resolves to the stage literally named for it (Biblical/Qumran
Aramaic). Sefaria’s Aramaic splits Talmud-side and Targum-side, the
Targumim carrying their sociolect. And eBL’s canonical literature —
98% first-millennium — now carries Standard Babylonian as a composed
register: akk:na/lit, a variety refined onto the stage the period
rules had already assigned, 7,240 documents.
Every query surface speaks --lect. Search had it; now parallels
filters its candidates by stage, define scopes dictionary shelves by
their resolution (the DÉRom shelf sits under Proto-Romance, never bare
Latin), cognates scopes witnesses of the asked anchor (“cognate sets
whose Greek witness is Koine”), and every MCP passage envelope carries
the resolved lect when the document has one.
No adapter without lects. Every passage-serving source in the
library now has a recorded lect posture — a rule, an override, or a
declaration: identity (the bare code is the whole claim — an honest,
recorded answer), pending (a named refinement candidate awaiting a
ruling), dates (per-document inference rides), codemap (universally
refined). The suite fails on a posture-less source, and a new
nabu lect suggest census tells an arriving source what could refine
it before its first document lands.
And the shelf keeps growing. The classical Aggadah core — Pesikta DeRav Kahana, the Tanchumas, Pirkei DeRabbi Eliezer, Seder Olam and their kin, thirteen titles chosen one by one where categories could not choose them — joins the Sefaria shelves for the next sync.
The registry itself stays deliberately pre-1.0: the layer proves itself as an instrument first; versions come when the vocabulary stops moving.