Ontology Validator
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A component in an agentic system that uses formal ontologies to check LLM outputs and proposed actions for domain correctness before they are executed or returned, sitting external to the LLM itself. Frank Coyle (UC Berkeley) advocates it as a principled guardrail mechanism grounded in Neurosymbolic AI.
Core Concept and Role
Coyle frames the Ontology Validator as a symbolic reasoning layer that operates alongside — not inside — the LLM, using structured domain knowledge to catch errors that natural-language prompt engineering cannot reliably prevent: "you can have a reasoner built on ontology to check keep the LLM on track". The validator is specifically described as "operating with these ontologies about our domain", meaning it reasons over domain-specific formal models rather than general heuristics.
Technical Composition
The validator is built from two W3C semantic web technologies:
- RDFS (RDF Schema) — for defining class hierarchies and domain/range constraints
- OWL (Web Ontology Language) — for richer logical constraints and automated reasoning
Both are cited explicitly as the "support technologies" underlying the guardrails mechanism: "for the guardrails, I'm referring to these concepts these support technologies with RDFS and owl".
Placement in the Agent Tool-Use Loop
Coyle prescribes a two-stage validation pattern within an agent's tool-use cycle:
- Input boundary — Use Pydantic to validate parameter types before tool calls: "check your types with Pydantic and then check your results with the ontology"
- Output boundary — Use the ontology reasoner to validate results after tool calls: "check your results with the ontology"
This places Pydantic as a syntactic/structural check and the Ontology Validator as the semantic/domain check, together bracketing the LLM's action.
No-Side-Effects Principle
A key design rule tied to the validator is that agents should produce no observable side effects until their proposed actions have passed ontology validation: "You want to run them through the ontology first and make sure that works". This treats the validator as a pre-commit gate, not a post-hoc audit.
Relationship to Neurosymbolic AI
Coyle situates the Ontology Validator within the broader Neurosymbolic AI paradigm, in which symbolic reasoning is used to constrain neural systems: "neuro-symbolic AI sort of represents a way to keep the LLM on its guardrails". The validator is thus an instantiation of that paradigm specifically applied to agentic tool use.
Points of Disagreement
No other speakers in the available material address the Ontology Validator directly. Coyle's position is presented without recorded opposition in these sources.