Context Graph Glossary
Every Key Term Defined
The authoritative reference for context graph and ontology terminology. 47 terms covering decision context graphs, decision ontology, ontology engineering, OWL, SHACL, PROV-O, pre-execution enforcement, accountable agents, MCP, and the full vocabulary of production AI agent infrastructure.
Context Graph
A decision-aware knowledge structure that determines what is valid, applicable, and authorized in a given situation, enabling AI agents to operate deterministically, auditably, and at enterprise scale.
A context graph extends knowledge graphs with five additional layers: temporal validity, applicability logic, exception handling, decision traceability, and provenance. Where a knowledge graph answers 'what is true?', a context graph answers 'what is valid right now, for this situation?' This makes it the decision infrastructure layer beneath autonomous AI systems.
Decision Context Graph
A hypergraph-based decision substrate that combines facts, relationships, rules, exceptions, and decision traces into a single queryable structure agents must consult before acting.
A decision context graph extends the context graph pattern with an explicit pre-execution requirement: every agent action is preceded by a graph traversal that validates applicability, scope, temporal validity, and policy compliance. Unlike memory or retrieval systems, it produces a deterministic pass or fail outcome rather than a probabilistic answer. It is the structural unit of accountable AI infrastructure used in production agent platforms such as Rippletide.
Pre-execution Enforcement
The architectural principle that policies, constraints, and validation logic are evaluated before an agent acts, not after the fact through monitoring or output filtering.
Pre-execution enforcement inverts the standard observability model. Rather than detecting violations after they occur (post-hoc logging, runtime monitoring, output filters), the system blocks invalid actions before execution. This is the difference between auditing a decision and preventing it. In production agent infrastructure, pre-execution enforcement requires a decision context graph that the agent must consult, not a sidecar that watches the agent.
Accountable Agent
An AI agent whose every action is preceded by validated context, constrained by enforceable policies, and recorded with a causal decision trace, by structure rather than by hope.
Where 'autonomous agent' emphasizes capability and independence, 'accountable agent' emphasizes governance and traceability. The distinction is structural: accountable agents are not built by adding monitoring to autonomous ones, but by constructing them on top of a decision substrate that enforces validation pre-execution and produces audit-grade traces post-execution. This is the architectural shift required to deploy agents in regulated, high-stakes environments.
Knowledge Graph
A structured representation of real-world entities and the relationships between them, storing facts as triples (subject -> predicate -> object).
Knowledge graphs model the world by mapping entities, relationships, and semantic meaning. Google's Knowledge Graph, Wikidata, and enterprise knowledge graphs are prominent examples. They excel at information retrieval, discovery, and semantic enrichment but do not govern decisions: they describe relationships without enforcing temporal validity, applicability, or policy constraints.
State Drift
A divergence between an AI agent's internal assumptions about the world and the actual state of the systems it interacts with.
State drift occurs when an agent's model of reality falls out of sync with execution state across tools and systems. This causes subtle reliability degradation as the system scales, not through dramatic failures, but through compounding minor errors. State drift is the primary reason production AI agents degrade over time, and it cannot be solved by observability alone.
Decision Trace
A recorded reasoning chain that captures what data was consulted, what alternatives were considered, what outcome resulted, and why that decision was justified.
Decision traces are what separate a context graph from a knowledge graph. They enable every decision to be replayed, audited, and queried. In production AI systems, decision traces provide the audit trail required for compliance, debugging, and trust. They capture not just the decision itself but the full chain of evidence and reasoning that led to it.
Causal Decision Trace
A complete reconstruction of why an agent made a decision, including which context was retrieved, which policies applied, which exceptions triggered, and how the final action was justified.
A causal decision trace differs from a log or audit trail in two ways: it captures causation (which inputs drove which outputs), not just chronology, and it is replayable. The same trace can be re-executed against historical context to verify the decision was correct under the rules in force at that time. Causal decision traces are produced natively by decision context graphs, not bolted on after the fact.
Decision Substrate
The persistent, queryable foundation an agent decides on top of, providing facts, relationships, rules, scope, and decision traces as a unified structure rather than scattered across systems.
A decision substrate is what a decision context graph provides at the architectural level: a single coherent layer that an agent's reasoning operates against, instead of multiple disconnected sources (databases, vector stores, configuration files, prompt instructions). Without a decision substrate, every agent decision recombines context ad hoc and produces inconsistent outcomes for similar inputs. With one, decisions are reproducible.
Temporal Validity
The constraint that rules, contracts, policies, and data have effective dates: expired logic cannot execute, and future-dated rules do not apply prematurely.
In enterprise systems, almost everything has a time dimension: contracts expire, policies get superseded, approvals have windows. Temporal validity makes time a structural constraint rather than optional metadata. A context graph with temporal validity can answer: 'Was this rule valid at the time the decision was made?' and 'Which version of this policy was in effect on date X?'
Applicability Logic
The mechanism that determines which rules, policies, or constraints apply to a given situation, based on context, conditions, and scope.
Not every rule applies to every situation. Applicability logic is the layer that evaluates context and determines which subset of rules, policies, and constraints are relevant. This is critical for enterprise AI agents that must navigate complex regulatory environments where different rules apply to different jurisdictions, customer segments, product types, or time periods.
Provenance
The origin, authority, confidence score, and data lineage of every piece of information in a context graph.
Provenance tracks where data came from, how trustworthy the source is, whether it has been corroborated, and how it was transformed. In a context graph, provenance is a first-class citizen embedded in every node and edge, not optional metadata stored elsewhere. This enables trust scoring, conflict resolution, and source attribution for AI-generated decisions.
Context Engineering
The practice of designing, structuring, and managing the contextual information that AI systems use to make decisions.
Context engineering goes beyond prompt engineering. While prompt engineering optimizes individual inputs to language models, context engineering designs the entire information architecture that surrounds AI decision-making, including what context is available, how it is structured, when it expires, and how conflicting contexts are resolved. It is the discipline of building reliable context graphs.
Context Collapse
The loss of critical contextual information as data flows through AI pipelines, transformations, and summarizations.
Context collapse happens when information is processed through multiple stages (retrieval, chunking, summarization, embedding) and the original reasoning, conditions, and qualifications are stripped away. By the time context reaches a decision point, the nuances that made it actionable are gone. Context graphs prevent this by preserving the full contextual chain.
Tool Misalignment
When an AI agent's actions through external tools do not align with the actual constraints, permissions, or intended outcomes of the task.
Tool misalignment occurs when an agent selects or parameterizes tools incorrectly, not because the LLM lacks capability, but because the constraints governing tool usage are not structurally encoded. The agent may have the right information but apply it to the wrong tool, with the wrong parameters, at the wrong time. This is an infrastructure problem, not an intelligence problem.
Decision Infrastructure
The structural layer between language models and execution that validates, constrains, and governs AI agent decisions before they are acted upon.
Decision infrastructure is the missing architectural layer in most production AI systems. It provides deterministic validation of state transitions, enforcement of business rules and policies, exception handling, and audit-grade traceability. A context graph serves as the foundation of decision infrastructure: it is the system that ensures AI agents make authorized decisions, not just plausible ones.
Deterministic Validation
The process of verifying that an AI agent's proposed action satisfies all applicable rules, constraints, and policies before execution, with a guaranteed pass/fail outcome.
Unlike probabilistic outputs from language models, deterministic validation produces binary results: an action is either valid or it is not. There is no 'probably allowed.' This is achieved by evaluating proposed actions against the context graph's rules, temporal constraints, applicability conditions, and exception logic. It is what makes enterprise AI agents trustworthy.
Exception Handling (Context Graph)
The treatment of exceptions, overrides, and edge cases as first-class citizens in a context graph, not as afterthoughts or bugs.
In enterprise reality, exceptions are the norm, not the edge case. A manager override, a regulatory exemption, a contractual carve-out: these are all exceptions that must be modeled explicitly. Context graphs treat exceptions as structured entities with their own provenance, temporal validity, and decision traces, rather than hard-coded special cases.
Entity Resolution
The process of identifying whether incoming data refers to an entity that already exists in the graph, merging duplicates, and flagging conflicts.
When new data enters a context graph, entity resolution determines whether it represents a new entity or an update to an existing one. Unlike simple deduplication, entity resolution in a context graph considers provenance, temporal context, and confidence scores. Conflicts between sources are flagged with metadata rather than silently overwritten.
Decision Intelligence
The discipline of capturing, analyzing, and leveraging organizational decision history to improve future decision-making.
Decision intelligence uses context graphs as institutional memory, preventing context loss when people leave and enabling evidence-based decisions. It goes beyond business intelligence (which analyzes what happened) by analyzing why decisions were made, what alternatives existed, and what outcomes resulted. Context graphs provide the structural foundation for decision intelligence systems.
AI Overview Optimization (AIO)
The practice of structuring content so that AI systems, including ChatGPT, Google AI Overviews, Perplexity, and Claude, can accurately understand, cite, and surface it in AI-generated answers.
AIO goes beyond traditional SEO by optimizing for how AI models understand and cite content. This includes providing structured data (JSON-LD), maintaining clear entity definitions, creating LLM-readable content formats (llms.txt), and establishing topical authority through comprehensive, well-structured content. Sites that structure content around context graph principles are more likely to be cited in AI-generated answers.
Retrieval-Augmented Generation (RAG)
A technique that retrieves relevant text chunks from a knowledge base and injects them into an LLM's prompt to ground its responses in specific data.
RAG improves LLM outputs by providing relevant context at inference time. However, RAG has fundamental limitations for production AI: semantic similarity does not equal applicability, chunks lose their surrounding context, there is no temporal awareness (expired information is retrieved alongside current), and there is no decision provenance. Context graphs address these gaps by providing structured, governed context rather than raw text retrieval.
Vector Database
A database optimized for storing and querying high-dimensional vector embeddings, enabling semantic similarity search across large datasets.
Vector databases power semantic search by converting text, images, and other data into numerical vectors and finding similar items based on distance metrics. They are essential for RAG pipelines and recommendation systems. However, they have no concept of temporal validity, decision traces, or applicability logic: they find what is similar, not what is valid or authorized.
Audit Trail
A chronological record of all decisions, actions, and state changes in a system, enabling post-hoc verification and compliance review.
In regulated industries, audit trails are mandatory. Context graphs produce audit trails naturally through decision traces and provenance tracking. Every decision is recorded with its inputs, reasoning, outcome, and the authority under which it was made. This makes context graphs essential infrastructure for AI systems operating in finance, healthcare, insurance, and other compliance-heavy environments.
Context Window
The maximum amount of text (measured in tokens) that a language model can process in a single inference call.
Context windows limit how much information an LLM can consider at once. Even with expanding context windows (100K+ tokens), the fundamental problem remains: stuffing more text into a prompt does not provide structured applicability, temporal validity, or decision governance. Context graphs solve a different problem than context windows: they ensure the right context is available, not just more context.
Guardrails
Output-level constraints that filter, validate, or modify an LLM's responses before they reach the user or downstream systems.
Guardrails constrain outputs: they check what the model said after it said it. Decision infrastructure (via context graphs) constrains inputs and governs the decision process itself. Guardrails are necessary but insufficient for production AI: they catch surface-level issues but cannot prevent state drift, ensure temporal validity, or provide decision traceability. Reliable AI systems need both.
Memory Degradation
The progressive loss of accuracy and relevance in an AI agent's stored memory as volume increases and context evolves over time.
As AI agents accumulate memory through sessions, conversations, and interactions, the quality of that memory degrades. Older memories may be outdated, contradictory to newer information, or irrelevant to current context. Without structured memory management (temporal validity, provenance tracking, and relevance scoring), agent memory becomes a liability rather than an asset at scale.
Structured Data (Schema.org / JSON-LD)
Machine-readable markup added to web pages that helps search engines and AI systems understand the content's meaning, relationships, and context.
Structured data using schema.org vocabularies (typically in JSON-LD format) tells search engines and AI models what entities exist on a page, how they relate, and what type of content is presented. For AIO, structured data is critical: it helps AI systems accurately categorize, understand, and cite content. Context graphs and structured data share a common principle: making implicit relationships explicit.
Model Context Protocol (MCP)
An open standard, governed by the Linux Foundation's Agentic AI Foundation, that defines how AI agents connect to external tools, data sources, and APIs through tools, resources, and prompts primitives.
MCP solves the transport and connection problem in agentic AI by standardizing how agents discover and invoke external capabilities. Created by Anthropic and donated to the AAIF in December 2025, it has 97M+ SDK downloads and is supported by every major AI lab. MCP delivers data; it does not validate, scope, or govern that data. Production agents typically combine MCP for transport with a context graph for governance.
Neuro-Symbolic AI
An architectural pattern combining neural models (LLMs, embeddings) with symbolic reasoning systems (rules, logic, constraint solvers) to achieve both generalization and deterministic enforcement.
Pure neural systems generalize but cannot guarantee correctness. Pure symbolic systems guarantee but cannot generalize. Neuro-symbolic AI splits the responsibility: neural components handle perception, generation, and unstructured reasoning; symbolic components handle validation, policy enforcement, and decision traceability. Decision context graphs are the symbolic substrate of production neuro-symbolic agents.
Policy-as-Code
The practice of expressing organizational policies, business rules, and compliance requirements as executable code that an enforcement layer can evaluate at decision time.
Policy-as-code transforms policies from documents read by humans into artifacts evaluated by systems. For AI agents, this is the only path to reliable enforcement: agents cannot interpret a 47-page compliance manual at decision speed, but they can be blocked by a context graph that has those policies encoded as queryable rules. Policy-as-code makes governance composable, version-controlled, and testable.
Scope Isolation
The structural enforcement that each agent or tenant operates within a defined namespace, with context, permissions, and decision rights that cannot leak across boundaries unless explicitly authorized.
In multi-agent or multi-tenant systems, scope isolation prevents one agent from seeing another's context, one tenant from accessing another's data, or one workflow from inheriting state from an unrelated execution. Without scope isolation enforced at the context graph level, multi-agent frameworks are one prompt-injection or one buggy handoff away from data leakage. Scope isolation is a precondition for enterprise-grade agent deployment.
Hallucination Prevention
The architectural strategy of preventing fabricated outputs by constraining what context an agent can act on, rather than detecting hallucinations after they occur.
Hallucination is a symptom of unconstrained context. An agent that can only act on context retrieved from a validated decision graph cannot fabricate a customer's entitlement, a policy clause, or a transaction state. Those must exist in the graph or the action is blocked. Hallucination prevention through structural constraint is fundamentally different from output filtering or confidence scoring, which only catch hallucinations after they have been produced.
Ontology
A formal specification of concepts, relationships, constraints, and rules in a domain, expressed in a way machines can reason over.
An ontology defines what exists in a domain (entities, classes, properties) and how those things relate (typed relationships, constraints, axioms). For AI agents, the ontology is the schema beneath the decision context graph: it determines what kinds of facts can be stored, what relationships are valid, and what conclusions can be inferred. Without an ontology, a context graph degrades into an unstructured pile of triples that cannot be queried with confidence.
Decision Ontology
An ontology purpose-built for governing agent decisions: it models not just entities and relationships, but applicability conditions, temporal validity, exception hierarchies, and decision provenance as first-class constructs.
Where a traditional domain ontology answers 'what is true in this domain?', a decision ontology answers 'what can an agent decide here, under which conditions, with which audit trail?'. It encodes policies as classes, exceptions as named relationships, scope boundaries as constraints, and decision outcomes as instances with provenance. Decision ontologies are the structural blueprint of decision context graphs used in accountable agent infrastructure.
Ontology Engineering
The discipline of designing, building, validating, and maintaining ontologies, including methodology, tooling, lifecycle management, and alignment with evolving domain knowledge.
Ontology engineering brings software engineering rigor to knowledge modeling: requirements gathering via competency questions, modular design, versioning, testing through SHACL constraints, and continuous alignment with reality. For production AI agents, ontology engineering is no longer optional. An agent operating on an ad-hoc, under-specified ontology produces decisions that cannot be audited or reproduced. The ontology must be a versioned, tested, owned artifact.
Competency Questions
Natural-language questions an ontology must be able to answer, used as both the design specification and the acceptance test for ontology engineering.
Before building an ontology, engineers write the questions it must answer: 'Which customers are eligible for this refund policy as of yesterday?', 'Which approvers override the default authorization for transactions over $10K?'. Each competency question maps to a SPARQL query that the finished ontology must support. For decision context graphs, competency questions are the unit of contract between policy owners and engineers. If the ontology cannot answer a competency question, the decision it informs cannot be trusted.
RDF (Resource Description Framework)
A W3C standard for representing information as triples (subject, predicate, object), forming the foundational data model of the semantic web and most ontology-based systems.
RDF triples encode every fact as a relationship between two entities (or an entity and a literal value). RDF is the lowest common denominator for knowledge graphs and ontology-backed systems. For decision context graphs, RDF provides the wire format, but is insufficient on its own: RDF expresses facts, not policies, applicability, or temporal validity. Those layers are added via OWL, SHACL, and decision-specific extensions.
OWL (Web Ontology Language)
A W3C standard for expressing rich ontologies with classes, properties, restrictions, and logical axioms, enabling automated reasoning over knowledge.
OWL extends RDF with the expressivity of description logic: subclass hierarchies, property restrictions, equivalence axioms, disjointness, and cardinality constraints. An OWL reasoner can infer new facts and detect inconsistencies. For decision context graphs, OWL is a useful but partial tool: it handles structural reasoning well, but does not natively express temporal validity or applicability scope. Production decision ontologies typically combine OWL for structure with custom annotations for decision-specific semantics.
SHACL (Shapes Constraint Language)
A W3C standard for validating RDF data against shapes (constraints on classes, properties, and values), enabling deterministic validation of knowledge graphs.
Where OWL is built around the open-world assumption (missing facts are unknown, not false), SHACL is built around the closed-world assumption (data must conform to declared shapes). For decision context graphs and pre-execution enforcement, SHACL is the workhorse: it produces pass-or-fail outcomes when an agent's proposed action is checked against the policy shapes. This determinism is what makes SHACL essential for accountable agent infrastructure.
PROV-O (Provenance Ontology)
A W3C standard ontology for representing provenance: who produced what, when, from what sources, using what process, and with what authority.
PROV-O provides the vocabulary for the provenance layer of any decision context graph. Every fact, decision, and derivation can be linked to its agent (who), activity (what process), and entity (what data), with timestamps and authority. For accountable agents, PROV-O is the foundation of causal decision traces. Instead of inventing custom provenance schemas, mature production systems align with PROV-O so that audit tools, regulators, and downstream consumers can interpret the trace without proprietary documentation.
Description Logic
A family of formal logics that underpin OWL and enable automated reasoning over classes, properties, and individuals with decidable inference procedures.
Description logics balance expressivity and computational tractability. They are the mathematical foundation that lets a reasoner derive new facts (subsumption, classification, satisfiability) from an ontology in bounded time. For decision context graphs, description logic provides the structural reasoning layer: which roles inherit which permissions, which exceptions override which defaults. Decision-specific concerns (time, applicability) typically sit outside DL and are enforced by SHACL or custom logic.
Closed-World Assumption (CWA)
The assumption that what is not asserted to be true is false. The opposite is the open-world assumption (OWA), where what is not asserted is unknown.
CWA and OWA produce radically different agent behavior. Under OWA (default in OWL), if the graph does not assert that a customer is authorized, the agent must treat authorization as unknown. Under CWA (default in SHACL and most enterprise systems), missing authorization is treated as denial. For pre-execution enforcement, CWA is the safe choice: the action is blocked unless the graph explicitly authorizes it. Mature decision ontologies make the CWA versus OWA boundary explicit per relationship, because the wrong assumption produces silent failures.
Ontology Alignment
The process of identifying correspondences (equivalence, subsumption, disjointness) between concepts in two or more ontologies, enabling interoperability across systems.
Real enterprises rarely have one ontology. They have a CRM ontology, an ERP ontology, a compliance ontology, partner ontologies. Ontology alignment is what lets a decision context graph integrate facts from these sources without flattening their semantics. For multi-agent systems, ontology alignment is the difference between agents speaking the same language by coincidence and agents speaking the same language by structural guarantee. Alignment is encoded as mappings with provenance, not embedded as silent translations.
Upper Ontology (Foundational Ontology)
A high-level ontology of generic concepts (entity, process, role, time, event) that domain ontologies extend, providing a shared philosophical and structural backbone.
Examples include BFO (Basic Formal Ontology), DOLCE, and SUMO. Upper ontologies answer foundational questions: is a contract an event or an object? is an approval a process or a state? For decision context graphs, anchoring a domain ontology in a credible upper ontology prevents reinvention and ensures that downstream extensions (temporal validity, provenance, applicability) compose cleanly. Skipping the upper ontology step is the most common reason enterprise ontologies fail to scale beyond a single use case.
Cite This Glossary
Joubert, P. (2026). “Context Graph Glossary.” The Context Graph. Retrieved from https://thecontextgraph.co/glossary
Related Resources
What is a Context Graph?
Complete definition, components, and enterprise use cases.
Context Graph vs Knowledge Graph
The architectural difference between mapping reality and governing decisions.
Context Graphs for AI Agents: Resource Guide
Curated resources, tools, and critical perspectives.
Building production AI agents? Missing a term?