Layers

Every document in the library carries its text and its language. Three add-on layers answer the other questions a reader brings to it: when was this written, where does it come from, and what kind of text is it? The layers share one doctrine:

• Extracted, never guessed — every claim is read from what the source actually ships (a date attribute, a findspot, a genre label), and what resists an honest parse stays visibly unclaimed.
• Honesty buckets are first-class — “undated”, “unmatched”, “unmapped” and upstream’s own “cannot determine” are counted answers, never silent gaps.
• Derived and rebuildable — all three layers regenerate from the canonical sources on every rebuild; the upstream claim is preserved verbatim beside every fold.
• Composable — each layer is a search filter, and they stack: nabu search "dis manibus" --kind funerary --from -100 --to 100 --place pleiades:423025.

When — the dates layer

As of 7 November 2026 — a live census: 2,465,111 documents carry dating bounds (of 2,854,118 live documents, 86%). nabu search --century -21, nabu list --by-date and nabu vocab --by-century all read the same per-document interval.

Every dating bound is an honest year interval extracted from whatever dating form upstream actually ships:

• Structured claims — EpiDoc origDate attributes, catalogue year columns, machine date attrs (the Korean chronicles’ per-entry 1617-01-00 dates), publication-year headers.
• Banded labels — Assyriological period labels (“Neo-Assyrian”) through a ruled period table; century-half grids (“15,2” = 15th c., 2nd half); anno-mundi annal years era-converted with the ambiguity kept as a span.
• Author dating — work-composition years and author life bands (Patrologia Latina), floruit strings (“fl. 892”).

What resists an honest parse stays raw and unbounded: prose datings (“Mitte 15. Jh.” without a grid), upstream’s own no-date fillers, circa strings with nothing firmer behind them. An interval is always a real range — no fake midpoints. Documents whose range spans several centuries are bucketed by their earliest bound below (1,094,147 such documents — the announced bias, not a hidden one).

The centuries

Where — the places layer

nabu place Girsu answers with the gazetteer card and every source’s holdings at that place. The layer has two halves: the gazetteers, held locally, and the matching decisions that connect a source’s verbatim place-name to an identity.

The gazetteers. The library never queries a gazetteer online — it holds them as canonical assets with provenance and derives one namespaced place index:

Held rows as of 16 October 2026.

Namespaces are parallel claims; equivalences between them are crosswalk data with provenance (3,438 rows: CIGS’s own columns + a Wikidata harvest), never inferred.

The decisions registry. nabu-places records the matching judgments — which identity a source’s verbatim place-name string denotes — each reviewable, in the pattern of nabu-lects. Three real rows tell the story: CDLI’s "Girsu (mod. Tello)" → matched cigs:GIR

• pleiades:912855; EDR’s "Mediolanum" — six Pleiades places carry that title, the row says the Insubrian Milan and names the five rejected homonyms; "Irisagrig (mod. uncertain)" → unlocatable — the site is unidentified in reality, and that is an answer, not a failure. An unlisted name is honestly unmatched; adapter-asserted upstream references always win over registry mints — the registry only ever fills silence.

Coverage. At the places program’s close (9 August 2026), 585,682 documents carried a machine place reference, of 708,905 that name a place at all — a 3.9× gain over the pre-program state; the largest sources sit at 74–87% matched (CDLI 250,484 of 337,572, Oracc 87%, EDR 86%, EDH’s 73,507 upstream-asserted). The maintained coverage detail lives in docs/places.md. Honest limits stay on the record: a dozen upstream references cite defective Pleiades ids (flagged loudly by the health invariants); long-tail names below the curated waves stay visibly unmatched until their wave lands.

What kind — the classification layer

As of 7 November 2026 — a live census: 2,309,369 documents carry a kind classification (80% of the kind-eligible library), across 27 class families. What kind of document is this — an epitaph, a receipt, a hymn, a school exercise? Each source ships its own genre vocabulary (EpiDoc inscription types, cuneiform catalogue genres, manuscript headings); those upstream labels fold onto one ruled cross-corpus class list (config/kind_classes.yml, with crosswalks to EAGLE, LCGFT and Getty AAT where those vocabularies recognize the class).

Classification is multi-label — a funerary poem is both — and the tree is optionally deep: a class is a family with named sub-classes (literary holds literary/poetry, literary/narrative, …), any prefix matches its whole family in queries, and a document upstream calls just “Literary” honestly carries the bare family head. The upstream claim is preserved verbatim beside the fold, so the ruled class never erases what the source actually said.

The classes

What stays honest

Two buckets are counted beside — never inside — the classes: unknown (47,058 documents) is upstream’s own “cannot determine”, a claim the library preserves rather than overwrites; unmapped (13,341 documents) holds values still awaiting a fold rule — a visible curation worklist, not a silent discard. Values reviewed and found not to be genre claims at all (a physical-layout tag, a copy marker) are declared as such in config and render as their own census section. Sources carrying nothing genre-shaped at all (133 sources, 544,749 documents) speak through their recorded postures, not through silence.

One query across the layers

nabu search --century -21                        # when
nabu place Girsu                                 # where: the card + holdings
nabu kind census                                 # what kind: the class board
nabu kind census --unmapped                      # the classification worklist
nabu search --kind literary/poetry --lang la     # a family narrowed to verse
nabu search lugal --place cigs:GIR --from -2200 --to -2000 --kind administrative

Every filter also rides the MCP tools (nabu_search, nabu_place, nabu_show) for conversational use, and every document card (nabu show) renders its date interval, place references and kind line side by side — the three layers on one card.