Data.
The files beside a collection.

A PDAI collection is a folder of documents and pages. Beside them stand DuckDB files, each written by one layer and read by others:

• the Stanza analysis — sentences, tokens, dependencies, features and entities, one row per token;
• the fact sidecar — letters, pages, sentences, persons, topics and footnotes as structured facts, together with the evidence space;
• the sentence record — sentences, clauses, keys and the normaliser’s footnotes with their grounds;
• the corpus dictionary of LWgrammar;
• the global register of individuals, shared by all collections.

A sentence is addressed by its document, its page and its number on the page; the number restarts on every page, so an address without the page is wrong. The same address appears in a footnote, in an explication, in a citation and in a deep link of the reader.

The certified document infrastructure that produces the pages and their analysis is described in PDAI (21 pages).

Collections spell names in many ways. The global register gives each person, place and institution one identifier — at present 13,081 persons, 2,892 places and 6,357 institutions. Collection builders write only the spellings they found; whether two spellings name one individual is decided by an agent following a written instruction, with a second reader, and an owner’s correction wins.

The normaliser reads the register and never writes to it. When a book names a person the register knows only from letters — Newton in Euler’s E57 — the binding is recorded as a candidate with its reasons, not entered as a decision.