The canonical usage guide for knowledge-stack terminology across the DreamLab mesh: which artefact the words taxonomy, ontology, knowledge graph, reasoning, grounding and context graph each name, the one-line gloss each term carries at first use on a public surface, and the conflations the mesh’s own copy must avoid. The source of truth is VisionFlow/docs/terminology.md; this page is its published mirror.
The stack, bottom-up
- Taxonomy — the is-a hierarchy, one part of the ontology and never a synonym for it. Ours is a lattice, not a tree: 1,401 classes declare more than one parent, and the overlap is published as data.
- Ontology — the formal vocabulary: OWL 2 classes, typed object properties and the axioms that constrain what a valid statement looks like. What can be said.
- Knowledge Graph — the ontology populated with live instances, at runtime: VisionClaw’s running graph, agents’ working graphs and pod-resident personal graphs assert individuals against these classes. What is said lives there. This site is the ontology in readable form — pure TBox, zero individuals by design.
- Reasoning — what follows, machine-checked at two points: the pipeline’s EL-profile reasoner gates the published closure at build, and the Whelk EL++ reasoner classifies the runtime graph before anything enters it. On these surfaces, “reasoned” always means machine-checked, with the check attributed to its owner.
- Context Graph — the assembly layer on top, consuming ontology, knowledge graph and operational sources to build an agent’s working set. Implemented here by the Ontology Loom.
- Semantic layer — the parallel concept from business intelligence: governed meaning for metrics over warehouse data. The mesh does not ship one, and the term is not used for any layer above.
Why this corpus is an ontology, not a knowledge graph
- Every page here declares an OWL class (pages and classes are 1:1); typed relations compile to restriction axioms; there are zero individuals, by design. That makes the corpus pure TBox — defensible as an ontology down to the last axiom, and an overclaim as a knowledge graph, which implies instance data it does not hold. When one word has to cover the published artefact, that word is the ontology (or corpus for the markdown source form). “Knowledge graph” points at the runtime layer, where individuals genuinely live.
The canonical sentence
- Researchers and agents write a corpus; a pipeline compiles it losslessly into a formal OWL 2 ontology, machine-checking every statement as it builds; at runtime the mesh’s knowledge graphs populate that ontology with instances under Whelk’s gate; and the Ontology Loom serves the checked ontology back into any model’s context at query time.
Usage rules
- The published artefact is the ontology; “knowledge graph” is reserved for the runtime layer where individuals live. Page and word counts belong to the corpus, the ontology’s source form.
- “Living ontology” is retired as a name for the whole artefact.
- “Reasoned” is earned, not decorative: unqualified reasoning is symbolic; LLM inference is always qualified as such.
- The provenance of the corpus is stated plainly in one line per surface — it is produced by an automated research process between researchers and agents, and it stands on its own merits — then the copy moves on.