← Labs Illustrative reconstruction · real NaPTAN stop types
JMR · Labs · Ontology

The same data, carved two ways. Watch what a retriever grounds on.

An ontology is not a diagram — it is an argument about how the world is carved up, and reasonable people disagree. Here is one machine-drafted modelling decision that no one reviewed. Flip it, and watch it silently rewrite the set a retrieval system would ground its answer on. Neither carve is wrong. Someone still has to sign it.

The decision
The carve — flip it

1 The instance that flips
is-a
drafted by the extractor reviewed by — no one signed by
2 What the retriever walks
walked · in the grounded setskipped · outside the carve
3 What a RAG answer grounds on
0
returned
extractive answer · assembled from the walked nodes only
This is a decision, not an answer. The lab records what you were willing to sign — and what you rejected.
Decision ledgerclear
Nothing signed yet. Every ruling — accept or reject — is recorded here.

An ontology is a decision

“What is the same, what is different, what is allowed to be true at once” is not read off the data — it is argued and written down precisely enough to execute. Each carve above is a defensible position, not a bug.

Retrieval is downstream of the carve

For RAG to ground well, the graph has to reflect real relationships, not surface co-occurrence. So a modelling choice made once, upstream, decides what a retriever can even reach — long before anyone reads the answer.

Generate, then sign

At government scale you bootstrap structure from content — the machine drafts. But a generated distinction can be confidently wrong, so curation is first-class: a human still has to sign the argument, or reject it.

Neither carve is wrong

The lab never names a winner. The point is that the choice has blast radius no one reviewed — read Ontologies Are Arguments and the GDS case study.

This is a faithful reconstruction of a mechanism, not live infrastructure: the stop types (RSE, TMU, BST, BCS) and the disputes are real NaPTAN modelling questions, but the retrieval walk and its two states are precomputed and hand-authored to be honestly wired end to end — there is no live model and the extractive answers are assembled from the walked nodes, never generated. The same judgment — an ontology is an argument, and retrieval is grounded on the argument you signed — is how I approached the real ontology generator for GDS.