Scale AI into systems that must be right.

LLMs generate probable answers. But when your system must be right - compliant, auditable, deterministic - you need a semantic backbone. VIStology builds ontology-based layers that ground AI in formal meaning and sound inference.

Neurosymbolic: LLM/Agents + Ontologies + Sound Inference Hallucination-resistant foundation

Neurosymbolic AI: the architecture for AI that must be right.

Talk to an expert

The AI/Agent boom created a new bottleneck: meaning

LLMs are impressive at generating and interpreting natural language, but enterprise and mission systems often require more than plausible text:

  • Teams need shared, consistent definitions across business and technical domains, tools and stakeholders
  • Systems must deliver unambiguous answers to complex, multi-source queries, no matter how multiple concurrent users phrase them.
  • Results must be reliable, auditable, and reproducible
  • Integration must survive changing requirements, protocols, and configurations

A semantic meta-layer that makes AI reliable

VIStology creates a formal, computer-processable semantic layer (ontologies + inference) that sits above your systems, data stores, and interfaces. It enables:

  • One shared vocabularies for requirements, specifications, and system descriptions
  • Cross-system reasoning with logically-consistent inference (implicit facts become explicit)
  • Complex semantic queries that remain stable even as data sources evolve
  • AI acceleration without AI uncertainty: use AI for efficiency, while the semantic layer remains the source of truth
AI boosts productivity. Ontologies protect correctness. Together: faster and safer.

Where each approach fits best

AI/LLM-first is great for:

  • Natural language interaction and summarization
  • Rapid exploration, drafting, clustering, extraction
  • User-facing copilots and conversational UX
  • Pattern recognition over large, messy corpora

Ontology + sound inference is essential for:

  • Deterministic answers with traceable logic
  • Integration across multiple systems and standards
  • Stable requirements/specs and shared definitions
  • Policy-based control and compliance verification
  • Multi-user querying that must be correct (not "likely")

The winning architecture is neurosymbolic: LLMs interface with humans; the semantic layer interfaces with real systems.

Building The Semantic Layer

1

Analyze

Examine existing system descriptions, requirements, and specifications to understand your domain.

2

Model

Capture domain knowledge as formal ontologies with precise, machine-processable semantics.

3

Connect

Link multiple systems into an interoperable semantic layer with shared meaning.

4

Query

Query across sources using SPARQL with consistent, shared vocabulary.

5

Infer

Derive implicit facts using sound reasoning engines for more complete answers.

Add new information types and new query types without re-architecting existing databases or integrated systems.

Why semantic technology matters more in an agentic world

Agents don't just answer questions, they take actions. That raises the bar:

  • Agents need ground truth constraints (what entities mean, what relationships are valid)
  • They need stable interfaces to enterprise knowledge, not brittle prompts
  • They need guardrails and policies that can be computed and verified
  • They need explainability for audit, safety, and stakeholder trust

Ontologies turn "prompting" into "programming meaning" and give agents a reliable map of the world.

The Extractor Changed. The Socket Didn’t.

Neurosymbolic AI is not a pivot for us. It is what VIStology has been building, under government contract, for more than two decades — and in that time the sub-symbolic half has been torn out and replaced in every era, while the thing it plugs into has not moved.

Follow the front end through the record. In 2002 it was a wavelet corner detector over simulated intensity and range imagery. In 2010 it was statistical text retrieval — tf·idf scoring, term-frequency-weighted proper-name vectors, link analysis — over a crawled foreign blogosphere. In 2013 it was regular expressions and a commercial geocoding API over half a million tweets. In 2022 it was a deep detection network on infrared video. Every one of those was the best available extractor of its era, and every one of them is now obsolete.

Now follow what each of them handed across the boundary. It is the same thing every time: typed assertions — individuals, the ontology classes they belong to, and the typed relations between them — expressed in a formal knowledge-representation language and handed to a sound reasoner. In 2002 that was Prolog facts and DAML annotation. From 2007 on it has been OWL/RDF individuals and property assertions, reasoned over by BaseVISor. The languages hardened. The contract did not change.

That is the whole point, and it is the reason this page is not a 2023 rebrand. Your LLM is this decade's extractor. It is a superb one — better at language than anything that came before it. It is also a component that plugs into a socket, and it inherits every failure mode its predecessors had: it is confident when it is wrong, it cannot tell you why, and it has no representation of the constraints your domain actually runs on. We have had the socket, and the reasoner behind it, since before the transformer was published.

The extractor changed. The socket didn't. Four generations of sub-symbolic front end stacked above one unchanging symbolic core. In 2002 a wavelet corner detector over simulated intensity and range imagery; in 2010 statistical information retrieval over a crawled blogosphere; in 2013 regular expressions and a geocoding API over half a million tweets; in 2022 the SOLO detection network emitting boxes and objectness confidence; and next, an open dashed slot for a large language or vision-language model. All but the 2010 system hand their output across the same interface — typed assertions naming individuals, their ontology classes, and the relations between them — into a formal ontology and a sound reasoner, which return a traceable, auditable answer. The ontology hardened from DAML to OWL 2 and the reasoner from Prolog to BaseVISor, but the interface did not change. The Extractor Changed. The Socket Didn’t. Two decades of sub-symbolic front ends — one enduring interface to formal, sound reasoning The extractor — replaced every era formal assertions ontology-typed individuals & relations 2002 Wavelet corner detector + a pixel-counting corner-typing heuristic OBTRA · simulated intensity + range imagery · Kokar & Wang, Northeastern (AFOSR / AFRL) — origin, pre-VIStology 2010 Statistical text retrieval — tf·idf scoring, proper-name vectors, link analysis IBlogs · 220,320 URLs crawled · relevance, specificity, timeliness, credibility · VIStology (AFOSR) 2013 Regular expressions carried by an OWL class + generic pattern matching + a geocoding API HADRian · 509,795 tweets, Boston Marathon · VIStology (AFRL SBIR Phase II) 2022 SOLO detection network — boxes + objectness confidence only, no class labels, ~10 FPS FLOCK · detector by Southwest Research Institute · semantic reasoner by VIStology (Army) next Your LLM / VLM the slot is open — the contract behind it is not The constant — the socket, and everything behind it The socket — the interface that has not changed typed assertions: individuals, the ontology classes they belong to, and the typed relations between them — in the ontology’s own vocabulary Prolog facts & DAML/RDF annotation (2002) → OWL/RDF individuals & property assertions (2007 – today) Formal ontology what a target, a place, a situation IS — declared, not coded DAML (2002) → OWL / OWL 2 (2007 – today) Sound reasoner derives what must follow — soundly, and traceably Prolog · SNARK (2002) → BaseVISor · OWL 2 RL · BVR rules (2007 – today) A traceable, auditable answer every conclusion traces back to the axiom or the rule that produced it The extractor is replaceable; the interface is not. That is why tomorrow’s LLM plugs into a contract fixed in 2002. BaseVISor enters the story in 2007, and has been the reasoner on the far side of the socket ever since.
Figure 1. Four sub-symbolic front ends, two decades of technology, one interface. The extractor has been a wavelet corner detector, a statistical retrieval stack, a bank of regular expressions, and a detection network — and tomorrow it can be a language model. What each hands across the boundary has not changed: typed assertions naming individuals, their ontology classes, and the relations between them. Behind the socket the ontology hardened from DAML to OWL 2, and the reasoner from Prolog to BaseVISor — but the contract they honor is the one written in 2002.

The record behind that figure splits into two lineages, each with its own page: the vision lineage, from the 2002 corner detector to FLOCK’s 2022 detection network, and the text lineage, from the 2010 blogosphere monitor to half a million tweets processed in 2013 — with, in every case, an exact account of what was measured and what we will not claim.

Use Cases

Scene Understanding

The vision lineage: from the 2002 paper that first drew the extractor–reasoner interface to FLOCK (2022), where a BaseVISor reasoner overrules a deep detection network with evidence it cannot represent — an object’s size in meters, and the terrain it sits on.

Learn more →

Social Media Situational Awareness

The text lineage: IBlogs measured an unfamiliar foreign blogosphere at corpus scale; HADRian mapped what half a million tweets were reporting, 11 minutes ahead of the newsrooms; and the trust work made NATO’s reliability ratings something a reasoner can revise.

Learn more →

Proof Points

Four funded programs, four different extractors, one architecture.

Vision

FLOCK — Detector Meets Reasoner

Army contract W909MY-19-P-0005, with Southwest Research Institute: a deep detection network behind a REST socket, a BaseVISor reasoner on the other side suppressing high-confidence false alarms with evidence the network cannot represent. The full record — and what it does not show — is on the Scene Understanding page.

Measured

HADRian — 11 Minutes Ahead of the Newsroom

AFRL SBIR contract FA8650-13-C-6381: from 509,795 tweets after the Boston Marathon bombing, 38 ranked places where people were reporting additional devices — every such report proved false, and the system never claimed otherwise — on average 11 minutes ahead of the CNN, New York Times and Boston Globe live blogs.

Scale

IBlogs — Statistical Front End, Corpus Scale

An AFOSR-funded monitor for an unfamiliar foreign blogosphere: 220,320 crawled URLs feeding four metrics of information value — the credibility metric validated against an independent panel of subject-matter experts — published in IEEE Internet Computing.

Trust

Reasoning Over Facts You Cannot Trust

US Army contract W15P7T-09-C-S031: NATO STANAG 2022 reliability and credibility ratings put into the ontology — reliability on the source, credibility on the proposition — and revised by BaseVISor rules as corroboration or conflict arrives.

Selected Publications

The peer-reviewed record behind two decades of neurosymbolic systems.

Improved Scene Understanding Through Semantic Reasoning and Online Learning

Moskal, J.J., Kokar, M.M. & Whittington, S.J.

Proc. SPIE 12122

View Publication

Metrics for Monitoring a Social-Political Blogosphere: A Malaysian Case Study

Ulicny, B., Matheus, C.J. & Kokar, M.M.

IEEE Internet Computing 14(2)

View Publication

Situational Awareness from Social Media

Ulicny, B., Moskal, J.J. & Kokar, M.M.

STIDS 2013

View PDF

Augmenting the Analyst via Situation-Dependent Reasoning with Trust-Annotated Facts

Ulicny, B., Powell, G., Brown, D., Kokar, M.M., Matheus, C.J. & Letkowski, J.

CogSIMA 2011

View Publication

The origin of the pattern predates the company. The paper below is not a VIStology program: it is AFOSR-funded research by Dr. Mieczyslaw Kokar — VIStology's president — then at Northeastern University. It reports no accuracy results and claims none. We cite it for one reason: it is where the interface between a sub-symbolic extractor and a formal reasoner was first drawn.

An Example of Using Ontologies and Symbolic Information in Automatic Target Recognition

Kokar, M.M. & Wang, J.

SPIE, Sensor Fusion VI, Vol. 4731

View PDF

See the full list of VIStology publications →

Related

Ontology DevelopmentThe knowledge-representation craft behind the models our reasoners run on. Semantic Technology IntegrationEmbedding ontologies, rules, and reasoning in the systems you already run. BaseVISorThe forward-chaining inference engine on the far side of the socket. PublicationsOur full body of peer-reviewed research.

Common Questions

LLMs generate probabilistic responses based on training data patterns - they produce text that's likely to be correct. Ontology-based reasoning uses formal logic to derive conclusions that are guaranteed to be correct given the defined rules and facts. When you need deterministic, auditable answers - especially for compliance, integration, or safety-critical systems - ontologies provide the foundation LLMs cannot.

Absolutely. Our neurosymbolic approach is designed to complement - not replace - your existing AI capabilities. We create a semantic layer that sits alongside your LLMs and ML models, providing them with ground truth constraints, validated inputs, and explainable guardrails. Your AI handles the "fuzzy" tasks; our ontologies handle the "must be right" tasks.

Most engagements start with a technical discovery session where we understand your domain, data sources, and integration requirements. From there, we typically scope a pilot project - often building a domain ontology and demonstrating semantic queries over your existing data. Successful pilots evolve into production integration, training, and ongoing support relationships.

We work with both. While much of our R&D experience comes from DoD contracts, our semantic technology expertise applies equally to commercial enterprises facing complex integration, compliance, or knowledge management challenges. Our products (BaseVISor, SCA-Pass) are available for commercial licensing.

Build AI systems that stay correct when everything changes

If you're building LLM copilots, agents, or decision automation over multiple systems, you need a semantic foundation that keeps results consistent, explainable, and integration-ready.

Schedule a technical consult