{
  "@context": {
    "@vocab": "http://schema.org/",
    "schema": "http://schema.org/",
    "faos": "https://example.org/faos#"
  },
  "@type": "ScholarlyArticle",
  "headline": "Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents",
  "author": [
    {
      "@type": "Person",
      "name": "Thanh Luong Tuan",
      "affiliation": {
        "@type": "CollegeOrUniversity",
        "name": "Golden Gate University",
        "address": "San Francisco"
      },
      "email": "tluongtuan@my.ggu.edu",
      "identifier": {
        "@type": "PropertyValue",
        "propertyID": "ORCID",
        "value": "0009-0000-1199-837X"
      }
    },
    {
      "@type": "Person",
      "name": "Abhijit Sanyal",
      "affiliation": {
        "@type": "Organization",
        "name": "Novartis Healthcare Pvt. Ltd.",
        "address": "Hyderabad, India"
      },
      "jobTitle": "Associate Director, Data, Digital & IT"
    }
  ],
  "datePublished": "2026-05",
  "version": "v3.0.1",
  "publisher": {
    "@type": "Organization",
    "name": "Foundation AgenticOS (FAOS)"
  },
  "keywords": [
    "Neurosymbolic AI",
    "Enterprise Ontology",
    "Large Language Models",
    "Agentic AI",
    "Domain-Driven Design",
    "Knowledge-Grounded Reasoning",
    "Multi-Agent Systems"
  ],
  "abstract": "Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Foundation AgenticOS (FAOS) platform that addresses these limitations through ontology-constrained neural reasoning. We introduce a three-layer ontological framework—Role, Domain, and Interaction ontologies—grounding LLM-based enterprise agents. We formalize asymmetric neurosymbolic coupling: current enterprise systems constrain agent inputs but not outputs, and propose mechanisms extending this coupling to output-side validation. A controlled experiment across five industries and three LLMs finds ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy and Role Consistency with large effect sizes. Improvements are greatest where LLM parametric knowledge is weakest, particularly in Vietnamese-localized domains.",
  "articleBody": "This paper presents a neurosymbolic architecture for enterprise AI agents using a three-layer ontology framework to ground LLM reasoning. It formalizes asymmetric neurosymbolic coupling and proposes extensions for output validation and ontology evolution. Empirical evaluation across 1,800 runs shows significant improvements in grounded agents, especially in low parametric knowledge domains.",
  "hasPart": [
    {
      "@type": "DefinedTermSet",
      "name": "Three-Layer Enterprise Ontology Framework",
      "description": "Defines enterprise ontology as a triple O = ⟨R, D, I⟩ where R is Role Ontology, D is Domain Ontology, and I is Interaction Ontology, each grounding agent behavior distinctly.",
      "hasDefinedTerm": [
        {
          "@type": "DefinedTerm",
          "name": "Role Ontology",
          "description": "Encodes how organizational roles think, decide, and communicate, including decision patterns, metrics focus, communication style, expertise domains, and approval authority."
        },
        {
          "@type": "DefinedTerm",
          "name": "Domain Ontology",
          "description": "Captures industry-specific concepts, relationships, metrics, and governance constraints, organized hierarchically by verticals."
        },
        {
          "@type": "DefinedTerm",
          "name": "Interaction Ontology",
          "description": "Formalizes organizational workflows as typed handoff patterns, approval chains, and escalation paths between roles."
        }
      ]
    },
    {
      "@type": "DefinedTermSet",
      "name": "Neurosymbolic Coupling Taxonomy",
      "description": "Classifies coupling of symbolic ontological knowledge with neural LLM reasoning into input-side, process-side, and output-side coupling.",
      "hasDefinedTerm": [
        {
          "@type": "DefinedTerm",
          "name": "Input-Side Coupling",
          "description": "Constrains LLM inputs via context injection, tool discovery filtering, and governance thresholds."
        },
        {
          "@type": "DefinedTerm",
          "name": "Process-Side Coupling",
          "description": "Constrains reasoning process through autonomy gates, quality judge verification, and escalation mechanisms."
        },
        {
          "@type": "DefinedTerm",
          "name": "Output-Side Coupling",
          "description": "Proposed mechanism to constrain LLM outputs against ontological definitions after generation."
        }
      ]
    },
    {
      "@type": "HowTo",
      "name": "Ontology-Constrained Context Resolution",
      "description": "Steps to generate ontologically grounded context for an enterprise AI agent.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Load Ontology",
          "text": "Load the industry-specific ontology for the tenant."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Merge Customizations",
          "text": "Merge tenant-specific overlays with the base ontology."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Extract Role Definition",
          "text": "Extract the role ontology definition for the agent's role."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Resolve Domain Context",
          "text": "Resolve domain ontology context relevant to the user query."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Resolve Interaction Context",
          "text": "Resolve interaction ontology context for the role."
        },
        {
          "@type": "HowToStep",
          "position": 6,
          "name": "Serialize Context",
          "text": "Serialize role, domain, and interaction contexts into raw context."
        },
        {
          "@type": "HowToStep",
          "position": 7,
          "name": "Optimize Context",
          "text": "Optimize serialized context with priority truncation (Role > Domain > Interaction) within token budget."
        },
        {
          "@type": "HowToStep",
          "position": 8,
          "name": "Return Context",
          "text": "Return the optimized ontologically grounded context for injection."
        }
      ]
    },
    {
      "@type": "HowTo",
      "name": "Ontology-Constrained Tool Discovery",
      "description": "Process to filter and rank tools for agents using ontological domain hierarchies.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Tag Skills with Domain Paths",
          "text": "Assign domain ontology paths to skills (e.g., fintech.payments.card_networks)."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Specify Domain Context",
          "text": "Discovery queries specify the domain context for tool discovery."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Calculate Semantic Score",
          "text": "Compute semantic relevance score using ts_rank between skill and query."
        },
        {
          "@type": "HowToStep",
          "position": 4,
          "name": "Calculate Ontological Domain Match",
          "text": "Score domain match based on exact or ancestor match in ontology hierarchy."
        },
        {
          "@type": "HowToStep",
          "position": 5,
          "name": "Calculate Capability and Role Match",
          "text": "Score capability and role relevance of skills to the query."
        },
        {
          "@type": "HowToStep",
          "position": 6,
          "name": "Filter by Governance Thresholds",
          "text": "Filter skills by quality thresholds specific to regulated domains."
        },
        {
          "@type": "HowToStep",
          "position": 7,
          "name": "Rank and Return Eligible Skills",
          "text": "Rank eligible skills by combined score and return for agent use."
        }
      ]
    },
    {
      "@type": "Question",
      "name": "What is asymmetric neurosymbolic coupling?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It is the observation that current enterprise AI systems constrain LLM inputs using ontological context and tool filtering but do not validate outputs against the same ontological definitions, allowing agents to produce outputs that violate constraints."
      }
    },
    {
      "@type": "Question",
      "name": "What are the three layers of the enterprise ontology model?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The three layers are Role Ontology (how roles reason and communicate), Domain Ontology (industry-specific concepts and metrics), and Interaction Ontology (organizational workflows and handoff patterns)."
      }
    },
    {
      "@type": "Question",
      "name": "How does the FAOS platform implement input-side neurosymbolic coupling?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Through context injection of ontological definitions into LLM prompts, ontology-constrained tool discovery filtering, and governance thresholds that filter skills by quality."
      }
    },
    {
      "@type": "Question",
      "name": "What is the inverse parametric knowledge effect?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "It is the phenomenon where ontological grounding provides the greatest benefit in domains where the LLM's parametric knowledge is weakest, and can sometimes reduce performance where parametric knowledge is strong."
      }
    },
    {
      "@type": "Question",
      "name": "What metrics were used to evaluate ontology grounding?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Terminological Fidelity, Metric Accuracy, Regulatory Compliance, and Role Consistency."
      }
    },
    {
      "@type": "Question",
      "name": "What were the main findings of the empirical evaluation?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Ontology-coupled agents significantly outperformed ungrounded agents on Metric Accuracy and Role Consistency across three LLMs, with the largest improvements in Vietnamese domains where parametric knowledge is sparse."
      }
    },
    {
      "@type": "Question",
      "name": "Why are ontologies preferred over just Retrieval-Augmented Generation (RAG)?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Ontologies provide structural constraints, composability, and verifiability that flat document retrieval cannot, enabling governance enforcement and output validation."
      }
    },
    {
      "@type": "Question",
      "name": "What are the proposed future extensions to the FAOS architecture?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Implementing output-side validation with OntologyValidator, closed-loop ontology evolution, adaptive context injection based on parametric knowledge, and full-scale human-expert evaluation."
      }
    },
    {
      "@type": "Question",
      "name": "What is the OntologyValidator?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "A proposed component to check LLM outputs against domain ontology constraints for terminological consistency, metric validity, interaction compliance, and governance alignment."
      }
    },
    {
      "@type": "Question",
      "name": "How does the FAOS platform handle multi-tenant ontology polymorphism?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "By supporting distinct domain instantiations per tenant using the same three-layer ontology schema, enabling composability and tenant-specific customizations."
      }
    },
    {
      "@type": "DefinedTerm",
      "name": "Metric Accuracy",
      "description": "The proportion of metric references in agent outputs that cite values within ontologically defined ranges."
    },
    {
      "@type": "DefinedTerm",
      "name": "Role Consistency",
      "description": "Alignment of agent output with the role's decision patterns, KPI focus, and communication style."
    },
    {
      "@type": "DefinedTerm",
      "name": "Terminological Fidelity",
      "description": "The proportion of domain terms in agent outputs correctly matching ontology definitions."
    },
    {
      "@type": "DefinedTerm",
      "name": "Regulatory Compliance",
      "description": "Correct citation of applicable regulatory frameworks in agent outputs."
    },
    {
      "@type": "DefinedTerm",
      "name": "Input-Side Coupling",
      "description": "Constraining LLM inputs using ontological context, tool filtering, and governance thresholds."
    },
    {
      "@type": "DefinedTerm",
      "name": "Process-Side Coupling",
      "description": "Constraining the reasoning process via autonomy gates, quality verification, and escalation."
    },
    {
      "@type": "DefinedTerm",
      "name": "Output-Side Coupling",
      "description": "Constraining LLM outputs post-generation against ontological constraints."
    },
    {
      "@type": "DefinedTerm",
      "name": "Ontology Evolution",
      "description": "The process of extending and refining the ontology based on agent experience and expert validation."
    },
    {
      "@type": "DefinedTerm",
      "name": "Governance Threshold",
      "description": "Quality gates that prevent low-quality skills from being available in regulated domains."
    },
    {
      "@type": "DefinedTerm",
      "name": "Semantic Skill Discovery",
      "description": "The process of filtering and ranking skills for agents using ontological domain hierarchies."
    },
    {
      "@type": "DefinedTerm",
      "name": "Asymmetric Neurosymbolic Coupling",
      "description": "The phenomenon where input constraints exist but output validation is lacking in enterprise AI systems."
    },
    {
      "@type": "HowTo",
      "name": "Implement Ontology-Constrained Neural Reasoning in Enterprise AI Agents",
      "description": "Steps to ground LLM reasoning using a three-layer ontology and neurosymbolic coupling.",
      "step": [
        {
          "@type": "HowToStep",
          "position": 1,
          "name": "Define Enterprise Ontology",
          "text": "Create Role, Domain, and Interaction ontologies capturing organizational roles, domain concepts, and workflows."
        },
        {
          "@type": "HowToStep",
          "position": 2,
          "name": "Implement Input-Side Coupling",
          "text": "Inject ontological context into LLM prompts, filter tools by domain and governance thresholds."
        },
        {
          "@type": "HowToStep",
          "position": 3,
          "name": "Extend to Output-Side Validation",
          "text": "Develop OntologyValidator to check LLM outputs against ontological constraints and enforce compliance."
        }
      ]
    }
  ],
  "image": {
    "@type": "ImageObject",
    "contentUrl": "https://arxiv.org/pdf/2604.00555v3.pdf#page=1",
    "caption": "Title page of the article with authors and abstract."
  },
  "citation": [
    {
      "@type": "CreativeWork",
      "name": "Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems",
      "author": [
        "Thanh Luong Tuan",
        "Abhijit Sanyal"
      ],
      "datePublished": "2026-05",
      "url": "https://arxiv.org/abs/2604.00555v3"
    }
  ]
}