Ask what kind of evidence settles a claim.

Send a sentence or a subject. Taxonomy returns the field that owns it, the evidence it must meet, how it would fail, and which other fields bear on it. Mathematics is the base. The same 984 identifiers are served as plain JSON and as SKOS.

Try it

A claim, a passage or a subject Add model suggestions (works only if the service has a model attached)

Endpoints

Call	Returns
`GET\|POST /v1/classify`	Owner field, cell, warrant, method, the evidence standard for that cell and what defeats it.
`POST /v1/audit`	Each sentence placed, plus flags where the wording claims more than the evidence type allows, and a count of claims by warrant.
`GET /v1/map?subject=`	The fields that bear on a subject, their cells, the question each field asks, and the evidence it would need. Also lists fields not expanded.
`GET /v1/view?subject=`	The lens question for all 25 fields.
`GET /v1/relate?a=&b=`	How strongly two fields or identifiers cite each other, and the fields they share.
`GET /v1/lookup/{id}`	One identifier: label, text, owner, citations in both directions.
`GET /v1/field/{code}.json`	A whole field: cells, objects, methods, warrants, bridges, scope and crosswalk.
`GET /v1/index.json`, `graph.json`, `skos.jsonld`	Fields and counts, the citation graph, the concept scheme.

curl -s -X POST $HOST/v1/classify -H 'content-type: application/json' \
  -d '{"text":"The importer must pay GST at the border under the Goods and Services Tax Act 1985."}'

curl -s "$HOST/v1/map?subject=AI-enabled+sunglasses"

How it works

Each claim is placed on four faces: the field whose evidence can establish it, the cell it concerns, the warrant it owes (proof, measurement, a dated record, a stipulation and so on, thirteen types) and the method behind it. The map endpoint runs a subject past every field and reports which ones answered. Tradition-internal authority (W13) is never accepted as evidence for a measurement, a date, a legal duty or the meaning of a word.

Model assist. Placement is done by fixed rules and gives the same answer every time. If the service has a model attached, assist=1 adds a second opinion. It can only choose from a fixed list of field codes, and it never replaces the rule-based answer.

Status

Working draft for expert review. The classifier is a keyword and rule engine written alongside the 50-claim pilot, so its pilot scores are a regression floor, not a measure of accuracy.
No human placement study has been run. Treat every placement as a suggestion.
audit flags are prompts for a reader, not findings about truth.
Indigenous knowledge (IK) and all Indigenous placements are provisional until Māori-led review.
map counts cue words. A field that does not appear has not been shown to be irrelevant.