The ontology argues back.
An extractor reads documents and proposes triples for the knowledge graph — exactly what an LLM does. Here a real description-logic reasoner is the guardrail: it admits what is consistent, derives what follows, and rejects what cannot be true — with a minimal proof. This is how a graph refuses to ingest a confident hallucination. It runs entirely in your browser; the reasoning is real, not scripted.
The proof is the product
Any engine can flag a clash. The work is the explanation: the minimal set of facts and axioms that force the contradiction, computed by removing facts until none is redundant — a true minimal justification, not a canned message.
Consistency is a guardrail
A retriever grounded on a graph is only as trustworthy as the graph. A reasoner at the write path rejects the triple that can't be true before it ever reaches retrieval — the check an LLM pipeline usually skips.
Written to execute
An ontology is an argument “written down precisely enough to execute.” Disjointness, functional roles, transitivity — these aren't documentation, they're a runtime that catches what prose review misses.
The rest of the pillar
See a modelling decision reshape retrieval in Grounded on What You Signed, the thesis in Ontologies Are Arguments, and the build in the GDS case study.
The reasoner is a real forward-chaining description-logic engine running on-device: it computes the deductive closure (subsumption, domain/range, transitivity), detects clashes (disjointness, functional roles, irreflexivity under a unique-name assumption), and extracts a minimal justification by black-box removal — standard ontology-debugging, not a per-example script. The T-Box and individuals are a deliberately small, real NaPTAN-flavoured slice chosen to produce genuine contradictions; a production ontology is far larger and uses a complete tableau reasoner. The judgment — an ontology is executable, and consistency belongs at the write path — is how I approached the ontology generator for GDS.