Search returns rankings. An ontology returns structure.
A neurosymbolic search engine claims a 2.7× accuracy win over e-commerce giants. The mechanism is the lesson: force the model to honor a graph, and you can audit what it did.
The headline is the accuracy number. The mechanism is the point: an ontology, not a corpus, is what makes a neurosymbolic model search precisely, because the model has to walk a versioned graph instead of free-associate a ranking.
A few weeks ago Onton announced a search model it says is 2.7 times more accurate than the world’s best e-commerce search engines. The accuracy claim will get tested by someone else; the mechanism deserves attention now. The model is neurosymbolic, which means its retrieval is constrained by an explicit ontology — products, categories, attributes, relations — instead of a similarity score over a corpus. A plain search engine returns a ranking and a hope. A neurosymbolic model returns a path through a graph, and a path is the only output you can inspect.
Search is a ranking, rankings are noise with manners
A lexical or vector retriever returns candidates ordered by a score. The score is a bet: items that look similar to the query, whether or not they are related. That is not an implementation detail, it is the definition — a ranking is a willing list of guesses, and each guess is a point where the system and the world can disagree silently. When the disagreement matters (a regulator asking what a product decision was based on), a score is not an answer.
An ontology is a constraint
DEINO’s Grounded Neutral Lattice is a four-layer ontology — Instruments, Actors, Companies, Governments — and querying it is not retrieval. The model walks the lattice; every edge it resolves is a versioned relation; every node it lands on is a versioned state. The cost is that structure must be maintained, and the return is that inference has a path. The GNL cannot free-associate across layers, because the layers are the constraint the model is forced to honor. That constraint is the whole of the reproducibility story.
A search engine returns what is similar. An ontology returns what is consistent — and consistency is the only thing you can replay.
Why structure is provable
- A ranking is a score and a hope: it cannot cite its work, because it had no path.
- An ontology is a versioned path: node A → edge → node B, each layer typed and each state hash-linked (BR-842).
- The manifest seals the path, not the score: that is the difference between something a buyer can read and something an examiner can check.
What structure does not do
An ontology bounds inference; it does not certify it. It constrains what can be produced, and it does not vouch for the result being correct — that is the calibration loop’s job. The ontology also has a maintenance cost: types must be defined, versioned, and retired, and a stale type is as misleading as a bad vector. We pay that cost on purpose, and we are honest that the layers are built but inert until calibration fills them. The constraint is the offer: a bounded inference that can be inspected, over a free inference you are asked to believe.
Similarity is a score. Consistency is a path. Regulators do not submit a score to court; they ask to walk the path.
— provenance note — the oracle