Foundational Research: Ontology-Native AI
Neuro-symbolic AI has a direction problem
Most work under the neuro-symbolic banner points one way: get symbolic knowledge into the neural model, as embeddings, as constraints, as a knowledge graph to attend over. The direction that gets far less attention is the reverse arrow, symbols checking what the neural model produced. For a language model authoring RDF, that reverse arrow is where the guarantee lives, and the cheapest instance of it catches an error no amount of knowledge injection prevents.
Two arrows, unequal attention
The Kautz taxonomy of neuro-symbolic systems, which Pascal Hitzler and others use to organise the field, distinguishes patterns by how the neural and symbolic parts compose: symbolic knowledge feeding the neural network, the neural network invoking symbolic routines, symbolic reasoning over neural outputs, and so on. The research energy has gone overwhelmingly into the feed-forward direction, teaching the network more symbolic structure. That is valuable, and it is where our own IES4 fine-tune lives. But a fine-tuned model is still a probabilistic device, and the experiments were blunt about the ceiling: on unfamiliar scenarios it kept emitting plausible-but-wrong terms. Injecting more symbols reduces that; it does not close it. Closing it needs the other arrow.
The reverse arrow, at its cheapest
Symbolic verification of neural output has a reputation for being heavy: theorem provers, full logical entailment, expensive reasoners. The closed-world vocabulary gate is the opposite. It is the smallest useful instance of the reverse arrow: take the triples the model produced, and check that every predicate and every rdf:type value is a term the ontology actually declares. No entailment, no reasoner, just set membership against a namespace the ontology owns. In our benchmark it caught 300 of 300 graphs carrying a fabricated term across three real vocabularies, with zero false positives, in milliseconds. It is the reverse arrow you can afford to run on every output, and it is a clean example of the pattern the taxonomy names but the field under-builds.
Why injection cannot substitute for verification
The tempting shortcut is to say a well-trained enough model would not need checking. But the two arrows guard against different failures. Knowledge injection shapes the distribution the model samples from; it makes the right term more likely. Verification makes a categorical statement about a specific output; it makes the wrong term impossible to commit. For anything built on a controlled vocabulary, defence data on IES, biomedical data on the OBO Foundry, the second guarantee is the one that matters, because a single fabricated identifier can quietly corrupt an integration that assumed every term was real. A more likely right answer is not the same as a guaranteed real one. This is the same complementarity we drew as boxes in the missing box: the generator proposes, the symbolic gate disposes.
"The neuro-symbolic conversation is mostly about pouring symbols into models. The under-built half is letting symbols check the model back. The cheapest version of that check, does this term exist, is also one of the most useful, and almost nobody runs it."
Fabio Rovai, Tesseract Academy
The benchmark, the method and the reproducible code are in the write-up. Existence is only the first thing the reverse arrow can check; the harder one, whether a real term is used in a sound place, is the subject of certified denotation.
Run the reverse arrow
Real pyshacl vs the closed-world gate, in your browser or in Colab.
