Working Glossary

Artificial general intelligence. A hypothetical system with broad, human-level capability across many domains, not narrow task performance. Discussed in the appendix as context, not as a near-term design target.

A category of system design that groups related approaches by where and how symbolic reasoning is integrated, for example differentiable logic, neural theorem proving, and LLM symbolic scaffolding.

The ability to inspect how a system produced an output, including rules fired, graph paths traversed, or intermediate decisions recorded. A core evaluation dimension in section 5.

A visible or reconstructable sequence of intermediate steps used to justify a model output. Distinct from chain-of-thought prompting; here the chain is an artifact, not a generation strategy.

An enterprise term for systems that combine multiple AI techniques, including machine learning, rules, optimization, and knowledge graphs. Often the framing under which NeSy enters Gartner-style industry analysis.

A training approach where logical or structural rules are converted into differentiable optimization objectives rather than enforced only at inference time. Central to the differentiable logic family.

Ensuring outputs obey predefined rules, logic, or invariants. The first dimension in the evaluation framework, E.1.

A knowledge graph built and maintained with explicit human or domain oversight rather than auto-extracted without review. Quality here directly affects retrieval reliability.

A neuro-symbolic framework combining neural networks with probabilistic logic programming, allowing learning and reasoning to operate jointly. An exemplar of the probabilistic logic programming family.

Methods that relax symbolic rules into continuous forms so they can be optimized through gradients. Architecture family 01 in this brief.

A change between training conditions and deployment conditions that degrades model performance. One of the key stresses that motivates adding symbolic structure.

A cognitive framing contrasting fast intuitive processing with slower deliberate reasoning. Used metaphorically in NeSy literature; the mapping is suggestive, not precise.

A learned numerical representation of data, words, entities, or concepts used to encode meaning for neural models. The primary interface between neural and symbolic components in many hybrid architectures.

The category of support behind a claim: peer-reviewed research, case study, vendor statement, or press coverage. Tracked explicitly in the evidence table.

Structured information represented as rules, ontologies, schemas, or graphs rather than implicit patterns encoded in model weights. The symbolic half of the NeSy equation.

The degree to which a system's behavior can be described in terms a human can understand. In NeSy, typically grounded in symbolic traces and auditable artifacts rather than post-hoc saliency maps.

A precise symbolic system for representing statements, rules, and inference with mathematically defined semantics. The classical foundation that differentiable and probabilistic NeSy methods relax or extend.

The policies, controls, documentation, and oversight practices used to manage AI risk and accountability. A primary driver of NeSy adoption in the LLM era.

A retrieval architecture that uses a knowledge graph as structured memory to improve grounded retrieval and multi-hop reasoning for LLMs. The most common commercial pattern today, and architecture family 05 in operational form.

A model output that is false, unsupported, internally inconsistent, or asserted with unjustified confidence. The engineering problem that makes NeSy feel urgent in production.

In regulatory contexts such as the EU AI Act, an AI system subject to stricter obligations because its use case can materially affect safety, rights, or critical decisions.

A governance requirement that people can monitor, intervene in, or review AI behavior where needed. One of the explicit obligations imposed on high-risk systems.

Creating or maintaining the rules, ontologies, schemas, and graphs a NeSy system depends on. Identified as one of the hardest operational costs: expensive to build, brittle when domains evolve.

A structured network of entities and relationships used to represent facts, support retrieval, and constrain reasoning. The symbolic backbone for GraphRAG and many scaffolding-based NeSy designs.

Large language model. A neural model trained on large text corpora to predict and generate language. The system that NeSy techniques increasingly aim to make more reliable.

A design pattern where an LLM is supported by external symbolic structures such as schemas, graphs, tool interfaces, planners, or verifiers. Architecture family 05 and the most-deployed NeSy pattern in 2026.

A NeSy approach combining first-order logic with differentiable learning so logical constraints can guide neural training directly through the loss function.

The property that outputs do not contradict stated rules, earlier claims, or required invariants. Tracked as evaluation dimension E.1.

Neural architectures that preserve interpretable logical structure while learning thresholds or rule strengths from data. Designed to keep the symbolic layer legible rather than opaque.

Absent or incomplete evidence that stresses purely neural systems and reveals the value of explicit reasoning structures. A key test condition for shift robustness.

A staged view of NeSy design where symbolic components enter during preprocessing, embedding, reasoning, or postprocessing. Frames the question of where the symbolic layer actually lives.

Short for neuro-symbolic AI. The umbrella term for systems combining neural learning with symbolic reasoning, and the subject of this brief.

The learned part of a hybrid system, typically responsible for pattern recognition, representation learning, or language generation. The fast side of the metaphor.

Methods blending theorem proving with neural representations, enabling differentiable or learned forms of logical inference. Architecture family 02.

AI that combines deep-learning methods with structured symbolic machinery such as logic, rules, programs, or knowledge graphs. See NeSy for the abbreviation used throughout the brief.

Out-of-distribution behavior. How a system performs when inputs differ meaningfully from its training data. A primary motivation for symbolic guardrails.

A formal representation of concepts and relationships within a domain. Often the structured backbone behind knowledge graphs, schemas, and constrained retrieval.

Facts or patterns stored implicitly inside model weights rather than retrieved from an external source. The complement of explicit knowledge, and the part NeSy aims to supplement.

A stage after generation where outputs are checked, corrected, constrained, or reformatted using symbolic logic or rules. One of the modular integration points.

A framework combining logical structure with uncertainty so rules and facts can carry probabilistic weight. The foundation for DeepProbLog and architecture family 03.

AI deployed in real operational settings where reliability, latency, governance, and maintainability matter, as opposed to research prototypes or benchmarks.

An approach where symbolic structure takes the form of executable programs, planners, or procedural logic. Architecture family 04, focused on systematic generalization rather than narrow memorization.

A scalability problem where possible reasoning paths grow too quickly for efficient inference. The primary computational bottleneck in neural theorem proving.

The practical force created by laws, standards, and compliance obligations that pushes organizations toward traceable, robust AI systems. Driver 01 in the brief.

The move from optimizing only for fluent generation toward optimizing for correctness, consistency, traceability, and governed deployment. The central thesis of the brief.

Fetching external information, documents, or graph nodes to ground model outputs in evidence. Necessary, but not sufficient, for verification.

The ability to remain reliable under perturbations, edge cases, missing evidence, or changing conditions. Evaluation dimension E.3.

A defined collection of symbolic constraints or logical statements used to guide, restrict, or verify system behavior.

A structured definition of allowed fields, types, or relationships used to constrain data and outputs. One of the lightest forms of symbolic structure, and often the first thing added to an LLM pipeline.

Retrieval that depends on organized entities, relations, or typed records rather than only free-text similarity. The mechanism behind GraphRAG's advantage over flat vector search.

The explicit reasoning layer in a hybrid system: rules, graphs, logic, planners, or verifiers. The slower side of the metaphor.

The collection of formal or semi-formal structures used for explicit reasoning: logic, rules, programs, and knowledge graphs. An umbrella term broader than any single technique.

Applying rules or verification after neural generation to improve consistency, validity, or compliance. The lowest-friction NeSy integration point, and one of the most commonly deployed.

Reasoning performed through explicit representations such as logic, rules, programs, or graphs rather than learned statistical patterns.

A shorthand metaphor for fast intuitive processing versus slower deliberate reasoning. Applied informally to describe the neural-symbolic split; the brief treats it as metaphor, not mechanism.

A formal reasoning process that derives conclusions from logical premises using valid inference rules. The classical technique softened by neural methods in architecture family 02.

A model action that invokes an external function, system, or API as part of a reasoning workflow. A lightweight form of symbolic scaffolding now standard in LLM agent architectures.

The ability to follow a system's output back through evidence, intermediate artifacts, and decision points. Evaluation dimension E.2 and a regulatory obligation for high-risk systems.

The mathematical goal optimized during model training. In NeSy, this may include data fit alongside constraint penalties or consistency losses, the constraint-as-loss pattern.

A component that checks whether an output satisfies rules, matches evidence, or meets logical and structural requirements. The presence or absence of a verifier is the key dividing line between reliable and merely symbolic-looking pipelines.

Intermediate actions or artifacts that can be independently checked rather than accepted as opaque model behavior. The practical unit of traceability in a deployed NeSy system.