Glossary of Terms
This glossary defines the terms the firm uses across its research. Each entry gives a plain definition and links to the paper where the idea is worked out in full. The terms come from the firm's own work, so the definitions here match how the firm applies them in engagements.
Data architecture and semantics
Dynamic context assembly
Dynamic context assembly builds the input a language model reads fresh for each request. It composes the smallest set of inputs sufficient for the one reasoning step the model is about to perform, drawn from a structured source the organization maintains for the purpose. A million-token window still loses attention as it fills, so the discipline is to assemble what the step needs rather than load everything and hope.
Defined in Dynamic Context Assembly.
Semantic layer
A semantic layer is anything that bridges what a model understands on its own and what the business needs it to understand. It earns its place only where the model's untutored reading would be wrong in a way that changes the answer. Where the model already reads a concept the way the business does, no layer is needed.
Defined in Semantic Layers in Enterprise Agent Systems.
The native test
The native test decides whether a concept needs a semantic layer. If the model would reach a correct answer from the question alone, no layer is needed for that concept. If its untutored reading would be wrong or incomplete, a layer is required.
Defined in Semantic Layers in Enterprise Agent Systems.
Constellation
A constellation is the set of concepts and relationships a single reasoning step needs held together at once. The unit of semantic design is the constellation a step requires, not a concept defined on its own. A layer worth building hands over the whole constellation in one piece rather than a stack of separate definitions the model has to reassemble.
Defined in Semantic Layers in Enterprise Agent Systems.
Reasoning unit
A reasoning unit is one discrete task the agent performs, such as validate an invoice, forecast cash, or explain a variance. The reasoning units come first and drive which constellations the semantic layer has to produce. The catalog of tasks the agent does precedes the catalog of concepts it needs.
Defined in Semantic Layers in Enterprise Agent Systems.
World-model layer classes
The world-model layer classes are the six kinds of domain knowledge a semantic layer supplies. Vocabulary maps user terms to entities in the data model. Structural places an entity in the dimensional context it sits in. Attribution ties a value to its source, its derivation, and its owner. Temporal resolves period, scenario, version, and fiscal calendar. Topology captures the dependencies between entities. Calculation captures the logic that produces a value.
Defined in Semantic Layers in Enterprise Agent Systems.
Cohesion mechanisms
Cohesion mechanisms are what keep the concepts in one constellation coherent with each other. A bounded context limits where a layer applies. Shared identity gives the concepts a common reference space. Closure checks confirm that every concept and relationship a step declares is defined before reasoning starts. Hierarchical containment carries meaning down the business's own trees where it exists naturally.
Defined in Semantic Layers in Enterprise Agent Systems.
Governance as middleware
Governance as middleware expresses control as deterministic functions invoked at every meaningful step of the traversal a financial agent performs. The reasoning layer explores and proposes. The governance functions evaluate each step and produce verdicts that decide what the agent encounters next. Control sits in the architecture rather than in instructions a prompt asks the model to follow.
Defined in Governance as Middleware.
Informational governance
Informational governance is the weaker pattern that places rules in the semantic layer, surfaces them to the agent as context, and trusts the reasoning layer to comply. It fails because the model is free to disregard what it was shown. The middleware position exists to replace it.
Defined in Governance as Middleware.
Attestation and certification
Attestation is the agent reporting that it considered the rules and chose to comply. Certification is the architecture demonstrating that the controls were applied, by recording the function invoked, the inputs it read, the verdict it produced, and the step at which it fired. Attestation states a belief about what happened. Certification proves it.
Defined in Governance as Middleware.
The agentic AI maturity curve
The agentic AI maturity curve is a five-stage read of where an enterprise AI program actually sits, from a single proof of concept through mature operation across many overlapping domains. The stages describe architectural reality rather than a project schedule. The recurring error is building for the stage the organization wants to be at rather than the one it is in.
Defined in Agentic AI and the AI Maturity Curve.
Navigation and execution
Navigation resolves an ambiguous input to a definite reference. Execution applies logic to a reference already resolved. The two modes pull in opposite directions as a program matures, because navigation across semantic layers gets easier as the data improves while execution inside a business process gets harder with each process the agent takes on.
Defined in Agentic AI and the AI Maturity Curve.
Stage-matching
Stage-matching is the discipline of building the architecture the current maturity stage calls for, not the one the organization aspires to. The right design at stage two is the wrong design at stage five, and forcing either onto the other is the common way these programs stall.
Defined in Agentic AI and the AI Maturity Curve.
Agent orchestration
The four-layer split
The four-layer split is the architecture that keeps a language model out of the orchestrator's seat. Orchestration runs as a deterministic workflow, intent detection and reasoning are handled by the model at narrow gates with curated context, and action runs as deterministic code. Orchestration and action stay deterministic, the model is asked one question at a time, and crossing those lines brings the hallucination problem back.
Defined in Where Agent Orchestration Breaks.
Discipline
A discipline is a body of knowledge with its own rules, vocabulary, and quality standards, maintained by people who specialize in it. A complex enterprise task often spans four at once, finance, software engineering, platform knowledge, and the client's own business. An orchestrator has to honor every discipline the task touches on every decision.
Defined in Where Agent Orchestration Breaks.
The five orchestration patterns
The five orchestration patterns are the distinct control-flow structures an agent system can take. A flat pattern is one model call. An iterative pattern is a fixed loop whose count the caller sets in advance. An agent-driven pattern lets the model choose the next step. A defined deterministic workflow runs a predefined sequence with explicit branches and gates. A recursive pattern organizes the work in layers and lets a failure propagate up to the layer that can handle it.
Defined in The Five Orchestration Patterns.
Recursive orchestration
Recursive orchestration organizes work in layers, each with a bounded scope, and assembles each layer's context from only what that layer needs. When a step deep in the stack fails, the failure is not retried in place. It propagates up to the layer that holds the context required to handle that kind of failure, which then re-invokes the work with more information.
Defined in The Five Orchestration Patterns.
Failure-driven collapse
Failure-driven collapse is the recovery rule inside recursive orchestration. A failure at one depth propagates up to the layer that can actually address it, rather than being retried where it happened. Fail at one level, recover at the level that holds the right context, and re-invoke with what was learned.
Defined in The Five Orchestration Patterns.
Deterministic-first
Deterministic-first design verifies every fact a rule can check before the model is ever called. The agent compiles what is known, identifies what is missing, and only then asks the model the narrow questions the rules could not answer. The model fills gaps and code does the work.
Defined in Enterprise Agents in Financial Systems.
The tool misconception
The tool misconception is the error of treating a language model's tool calls as reliable deterministic actions. A tool in a model framework is a description and a calling convention, and the model can call it wrong, skip it, or misread the result. Any step that touches money has to treat the tool call as something to verify, not trust.
Defined in Enterprise Agents in Financial Systems.
Risk-rated orchestration
Risk-rated orchestration builds risk into the workflow as an architectural property rather than a policy added later. Risk is classified early, often as the first deterministic check, and the verification and human approval a step gets are set by the risk it carries.
Defined in Enterprise Agents in Financial Systems.
Self-learning in enterprise agents
Self-learning agent
A self-learning agent changes its own behavior from what it sees in production, between runs, without an engineer editing it. What it changes is its working state, the instructions it gives the model, the context it assembles, and the rules it has saved. It never changes the code that runs it.
Defined in Self-Learning Architectures for Enterprise Agents.
The three categories of learning
Everything a self-learning agent can change falls into three categories, and none is safe to let the agent adjust on its own. Deterministic routing and workflows are the most contained. Prompt and context come next. Code is the one the agent never touches. What separates the categories is how far the damage spreads when a lesson is wrong.
Defined in Self-Learning Architectures for Enterprise Agents.
Learned rule as configuration
A learned rule is configuration, not code. It is a discrete entry that deterministic code reads before the agent acts and writes to only between runs. Keeping learning as data a human can inspect is what makes a self-learning agent governable.
Defined in Self-Learning Configuration in Enterprise Agents.
Rules, patterns, and antipatterns
Rules, patterns, and antipatterns are the three things a learned store holds. A rule pairs a condition with an action and runs identically on every case it matches. A pattern is a recurring case the agent has learned to recognize but decides nothing on its own. An antipattern is a case the agent has learned to refuse.
Defined in Self-Learning Configuration in Enterprise Agents.
Inheritance and override
A learned rule attaches at a level in the business's own hierarchy, such as an entity tree or a chart of accounts, and applies to everything beneath it. A more specific rule lower down overrides what a node would otherwise inherit. An empty entry turns an inherited rule off on purpose.
Defined in Self-Learning Configuration in Enterprise Agents.
The cadence of the close
The cadence of the close is the rhythm a learned rule is admitted on. A rule enters the store no faster than the cycle that can prove it belongs, and it moves up the hierarchy only after the close that validates it has actually happened.
Defined in Self-Learning Configuration in Enterprise Agents.
Learning instead of fixing
Learning instead of fixing is the trap where an agent quietly encodes a workaround for a broken process instead of surfacing the problem. Most of what an agent could learn is a deferred fix for process, data, or technical debt dressed up as a rule. A confidence score measures whether a workaround works, not whether it should exist at all.
Defined in Learning Instead of Fixing.
Process, data, and technical debt
Process, data, and technical debt are the three kinds of deferred work a learned rule usually stands in for. Process debt works around a manual or broken process. Data debt reconciles the same fact arriving in inconsistent forms. Technical debt covers a missing integration or a platform gap. Read together, the rule store maps where the organization still owes work.
Defined in Learning Instead of Fixing.
LLM architecture and training
The REINFORCE Sidecar
The REINFORCE Sidecar is a small reinforcement-learning network that runs alongside a model during training and learns the values of training knobs online from the model's own loss signal. It reads from the main model and writes back to the training-loop controls without taking part in the model's forward computation. In the headline result it beat an otherwise identical hand-tuned baseline by 42 percent at matched compute.
Defined in Reward-Driven Training Control.
The Fire-Aim-Fire loop
Fire-Aim-Fire is the iteration loop the firm uses to judge a feature. Fire builds a minimum version on a fresh fork. Aim adds the observability needed to tell whether the feature helps. Fire again keeps it, kills it, or adjusts it on what the telemetry says. The obligation the loop enforces is that no feature counts as useful until it is instrumented deeply enough for the data to render a verdict.
Defined in Research Posture.
Knob-ownership migration
Knob-ownership migration is the practice of moving each hand-picked number in a training stack onto a learned signal over time. A hand-set value is treated as temporary scaffolding around a reward path that has not been built yet. The reverse move, taking a decision back from the model and handing it to a human, has happened zero times.
Defined in Research Posture.
Layer roles
Layer roles are a four-part account of what a layer does in a trained transformer, computable from a finished checkpoint plus a short gradient trace. A hub routes information widely. A damper holds the residual stream in check. A passthrough contributes a near-identity, broad and small. A specialist memorizes specific patterns and fires hard only when it sees one.
Defined in Emergent Layer Roles and Functional Specialization.
Saturated positions
Saturated positions are feed-forward positions whose activation runs far above the layer average, past three standard deviations of the mean. Counting them is one of the measurements that classifies a layer's role.
Defined in Emergent Layer Roles and Functional Specialization.
Bridges
A bridge is a small learnable projection that connects two neighboring layers of different widths. The upstream layer writes at its own width, the bridge maps that to the downstream width, and the downstream layer reads at its own. Bridges are what let a model discover its own per-layer proportions during training.
Defined in Self-Discovering Architectures.
Getting started with AI and enterprise agents
AI as a sister enablement function
AI belongs beside IT as a peer enablement function, not inside it. IT concentrates infrastructure, security, and systems, while the AI function concentrates the skills that make those systems intelligent. Where AI reports matters less than getting the ownership split right and having IT enable the work rather than gate it.
Defined in Is AI Part of IT?.
The four responsibilities
A serious AI deployment carries four responsibilities, and no single function holds more than two. Architecture belongs to IT. Functionality and the meaning of the data belong to the business function. Agent development belongs to the AI function. Risk, compliance, and cost cut across all of them under a shared standard.
Defined in Is AI Part of IT?.
AI-native and AI-attached
An AI-native system is designed around AI from the start. An AI-attached system bolts AI onto a design that predates it. The destination for enterprise systems is AI-native, where the intelligence is part of the architecture rather than a layer added on top.
Defined in Is AI Part of IT?.
Make it right, then make it fast
Make it right, then make it fast puts correctness before optimization in every deployment. An agent that is fast and fragile, with no verification between it and the ledger, is worse than no agent. Optimization is the second job, and orchestration is where most of the cost is later won or lost.
Defined in I Have Working AI Agents. What Do I Do Next?.
Independent liaison
An independent liaison is an outside advisor a buyer brings in to guide a larger implementer and hold the buyer's interest. The most valuable work happens during selection and contracting, where the risks are identified before they are locked into a contract.
Defined in Your Board Wants AI. Where Do You Start?.
© 2026 Proforma Global. All rights reserved. Permission requests: [email protected].