Foundational Research: Ontology-Native AI
The missing box: where symbol-existence checking belongs
When a large language model writes RDF, it is one component in a hybrid system: a neural generator feeding a symbolic store. The value of thinking about these systems as explicit assemblies of boxes, as Frank van Harmelen and colleagues argue in their boxology of hybrid learning-and-reasoning, is that it makes you name the box you left out. Most LLM-to-knowledge-graph pipelines left out the same one.
Boxes, and the one that is missing
The boxology is a design-pattern language for hybrid systems: draw the components (a learned model, a symbolic reasoner, a knowledge base) and the arrows between them, and you can reason about what the whole system does and does not guarantee. Apply it to the common generative-KG pipeline and the picture is short. One box generates candidate triples from text. One box, SHACL, validates structure. An arrow runs from the second into the triple store. The pipeline assumes those two boxes cover the job. But there is a step between them that neither performs: checking that the symbols the neural box emitted actually refer to entities the symbolic world defines. That is symbol grounding, and it is a box of its own. Leave it out and a fabricated term flows from generator to store unchallenged, because the generator does not know it invented anything and the validator is not looking at existence.
Why the neural box cannot self-check
It is tempting to push grounding back into the generator: fine-tune it hard enough and it will stop inventing terms. This helps, and we have measured how much. Our IES4 language model cut the hallucinated-term rate from 0.937 to 0.010, and a validate-and-repair loop lifted in-distribution conformance to 100%. But the same experiments showed the ceiling: on scenarios it had not seen, the model kept inventing plausible-but-wrong spellings, and a repair loop that only says "that term is invalid" stalled where the model could not guess the correction. The generator, however good, is a probabilistic box. It cannot offer a guarantee about the closed set of real terms, because guarantees about closed sets are not what neural boxes produce. That is precisely the work a symbolic box exists to do.
What the box actually is
The symbol-existence box is small and deterministic. It sits on the arrow from generator to store, and it asks one closed-world question of every triple: is each predicate, and each rdf:type value, a term the loaded ontology declares? It polices only the namespaces the ontology owns, so it never touches standard vocabulary or your instance data, and its false-positive rate on clean graphs is zero. In our benchmark this box caught every fabricated term across three real vocabularies, 300 graphs, 418 fakes, where open-world SHACL caught none. The lesson is not that SHACL is wrong; SHACL is a different box doing a different job. The lesson is that a hybrid system with a hole where its grounding box should be will look, from the outside, exactly like a system that works, right up until the fabricated data is load-bearing.
"Draw your generative-KG pipeline as boxes and arrows, and the missing component is obvious: nothing checks that the symbols are real. Add that one small, deterministic box and the neural generator is finally safe to trust with a knowledge graph."
Fabio Rovai, Tesseract Academy
The generator and the grounding box are complementary, and neither is asked to do the other's job: the model removes the blank page, the symbolic gate guarantees the terms exist. That division is the whole architecture. Read the measured result, or see how the same split plays out inside a model in the IES4 study.
The grounding box, open source
The vocabulary gate and the full benchmark are in the open-ontologies engine.
