NLIT
◆ DEEP DIVE
Routing · document class → capability

How each document is read

the routing map as the manifest declares it: document class (printed text, table, handwritten, handwritten-locate, image/artwork, photo, map, corpus query) → capability → provider, with an egress marker showing which legs are local (lexical retrieval and embeddings) and which call an external API; distinguish configured routes from evidence-specific execution provenance

Documents do not read themselves, and they are not all read the same way. A document is often first turned into a plain, structured form that a machine can work through without tripping over the original layout. After that, the kind of document decides how it is read.

Printed pages and tables are transcribed automatically. Photographs, drawings and maps are described automatically, so that there is a record of what they show, including any words visible in them. Neither result is treated as final or complete. Wherever a claim leans on an automatic transcription, it is checked against the scan. An automatic description is first of all a finding aid: the people, places, dates and text it reports must be confirmed against the source before they are used, because a confident machine can be confidently wrong.

Handwriting is held to a stricter rule. An automatic pass may find where on a page the hand of interest sits, but no value that a claim will rest on is ever accepted from a single automatic reading of a hand. Such values go through the independent double reading set out alongside . Two readers of different makes read the crop cold. One of them is the project's own, and it then weighs both answers against the scan and settles the value.

The last kind of reading searches the whole collection for documents that bear on a question. It merges two rankings, one for matches to the words of the query and the other for matches to its meaning. Both run locally under the declared retrieval route. If the semantic index is out of date, the search must fail closed instead of quietly handing back old rankings.

Documents are read by outside services, not by a machine in the building: printed transcription and picture description happen elsewhere, and handwriting cross-reads use more than one reader, at least one of them an outside service. Searching the collection, by contrast, runs entirely on the project's own machines, both the word matching and the semantic embeddings. Where each reading took place, including whether material was sent out or stayed here, is disclosed and counts in judging it.

The manifest also lists second-opinion, image-upscaling and image-generation capabilities, but marks them as wired, not built into any workflow. That they appear in the configuration does not show that they were ever used on a source or on anything accepted.