classify-and-route

Stage 3 of the seven-stage pipeline. You decide what each digitized work is as a form (photograph, document, artwork, object, audio, video, or a compound of these) and route it to the target schema and prompt set that fit that form. The decision is made from deterministic signals first, with a small classifier reserved for the cases the rules cannot settle, and every step is readable and adjustable by the institution.

This stage sits between pre-processing and generation:

[2] Pre-processing  →  [3] Classification & routing  →  [4] Generation (LLM)
    derivatives,           detect content type,            first-pass record against
    OCR, transcription,    pick target schema +            the routed schema + prompt set
    format metadata        prompt set

See ../../docs/architecture.md for the full pipeline and ../../docs/the-improvement-loop.md for how routing changes are classified as generalizable improvements or local overrides.

What it reads and what it sets

Reads:

• The institution's application-profile.json (conforms to ../../schemas/application-profile.schema.json). Its target_schema and controlled_vocabularies are the institutional default. This stage refines that default per item.
• Stage 1 intake evidence, which rides on the record as collection_context.source_metadata in ../../schemas/metadata-record.schema.json: filename, folder path, accession notes, partial legacy records.
• Stage 2 pre-processing output: MIME type, file extension, technical/format metadata (dimensions, duration, page count), and the presence of OCR text or an AV transcript.

Sets:

• target_schema on the metadata record. The record schema is closed, so target_schema is the one field this stage writes. It carries the routed schema string (for example Dublin Core + VRA Core, PBCore, or METS (compound) + per-constituent schemas) into stage 4.
• The stage-4 prompt set: the base prompt ../../prompts/system/metadata-generation.v1.md plus the matching modality overlay under ../../prompts/modalities/ (for example a PBCore overlay for audiovisual works). The overlay adds the technical fields and the modality-specific guidance; the base descriptive logic is unchanged.

Do not invent a routing field or a classification-confidence field on the record. The record schema is additionalProperties: false and will reject one. When you want to record why an item routed the way it did, write it to the run manifest's notes (../../schemas/run-manifest.schema.json), not to the record.

The routing table

These are the default rules. They are deliberately in the open so a cataloguer can read exactly why an item was routed and described the way it was, and can adjust the rules without a vendor in the loop. There is no opaque service call here.

The schema choices follow the metadata-standards reference: Dublin Core is the interoperable baseline, VRA Core adds the Work/Image distinction visual works need, PBCore carries audiovisual technical and descriptive fields, MODS suits bibliographic-granularity textual material, and METS is the structural wrapper for compound objects. If a target schema you need is not in the table, add it to your overrides/ routing rules and contribute the detection improvement upstream if it is generalizable.

Rules first, classifier second

Resolve content type in this order, stopping at the first confident answer:

1. Deterministic signals. MIME type and extension from stage 2 settle most items: an audio MIME routes to PBCore, a video MIME routes to PBCore, a multi-page or manifest-wrapped object routes to compound. A dominant OCR text layer distinguishes a document from a photograph of text.
2. Source-metadata hints. Filename, folder path, and accession notes from stage 1 disambiguate the visual cases the MIME type cannot, for example separating a photograph from a digitized artwork or a three-dimensional object.
3. Small classifier, for the residue only. When the rules leave an item ambiguous, a small, inspectable classifier proposes a content type. Model is a parameter here as everywhere: a lightweight local model is appropriate, since this is a coarse form decision rather than the descriptive generation in stage 4. The classifier proposes; the rules and the human dispose.

Never-fabricate extends to routing. An item the rules cannot settle and the classifier rates low-confidence is surfaced for a human routing decision rather than routed on a confident guess. A wrong route sends the wrong prompt set to stage 4 and corrupts the record before review sees it, so a held routing decision is cheaper than a confident error.

Classify by form

This stage is where hard rule 4 is enforced first. Classify a work by the form of the digitized object rather than by what the object depicts. A photograph of a building is a photograph (still image, routed to Dublin Core + VRA Core); the building is a depicted subject that stage 4 describes with AAT and TGN, and it is never the basis for the route. A scanned letter about a sculpture routes as a document, since the form is textual even when the subject is an object. Getting this wrong at stage 3 is the most consequential routing error available, because it sends the entire work to the wrong schema and prompt set.

Compound objects

A compound object is a structural whole made of constituents (a multi-page diary, a folder of related images, an oral-history recording with a companion transcript document). Route it in two parts:

• The whole gets a structural wrapper, METS by default, which links the constituents and carries the object-level descriptive record.
• Each constituent routes through this same table on its own form, so the pages of a diary route as documents and a photograph inside a mixed folder routes as a photograph.

Preserve the parent relationship on every constituent record via collection_context.is_part_of so stage 4 describes each piece as part of the whole rather than as an atomized item.

Local rules versus upstream contributions

Institution-specific routing rules stay in the adjacent overrides/ workspace: a rule that one folder is always maps, or that a given accession is entirely glass-plate negatives, encodes one collection's layout and should not ship to other institutions. Generalizable detection improvements, for example better discrimination between a digitized document and a photograph of text, return to the shared repo. See ../../docs/the-improvement-loop.md and ../../CONTRIBUTING.md for the test and the process.

Usage

This is the cited audiovisual case (sample user prompt 6). The routing decision is the first thing that happens: detecting audio routes the collection to PBCore and prepends the local transcription stage, so generation reads the transcript rather than attempting vision on the audio.

This collection is digitized oral-history audio (./collections/oral-histories/).
Run the transcription stage locally first (Whisper, nothing leaves this machine),
then generate metadata using the PBCore target schema instead of Dublin Core, pulling
description and subjects from the transcript. Flag any named people for LCNAF verification.

Here the audio MIME type settles the content type by rule, the route sets target_schema to PBCore and selects the modalities/av overlay, and the dependency on a stage 2 transcript is why the transcription step runs before generation. The named-people flag is handled downstream by hard rule 5 in stage 4 and confirmed by the deterministic authority stage; this stage's only job is the form decision and the route it implies.

A second, classify-specific example, for a mixed collection where the rules need to run per item:

Read the manifest at ./runs/2026-week-of-civil-war/manifest.json. For each item, classify
its content type from the MIME type, the OCR-text presence, and the folder path, then set
target_schema and the stage-4 prompt set per the default routing table. Where an item is
ambiguous, list it with the competing types and let me decide rather than routing it
automatically. Write the routing rationale for the ambiguous items to the run manifest notes.

Output of this stage

For each work this stage produces:

• target_schema set on the record (conforming to ../../schemas/metadata-record.schema.json).
• The selected stage-4 prompt set (base prompt plus modality overlay), passed to generation.
• An optional routing rationale for ambiguous or human-decided items, recorded in the run manifest's notes (../../schemas/run-manifest.schema.json).

Items held for a human routing decision do not advance to stage 4 until the route is set.