Skip to main content

3. Key Definitions and Terminology

This section establishes a shared vocabulary for the architecture. These definitions are normative and are used consistently throughout the Architecture Definition Document. Where common industry terms are used, they are intentionally scoped to avoid ambiguity.

3.1 Capability​

A capability is:

A discrete, outcome-focused function that accepts defined inputs and produces a defined result, independent of implementation, execution mechanism, or consuming system.

Key characteristics:

  • Describes what outcome is produced, not how it is achieved
  • Is stable over time, even as implementations change
  • Can be executed via deterministic logic, AI-assisted processes, or hybrid approaches
  • Is reusable across multiple systems and contexts

A capability is defined by its contract, not by the service or component that implements it.

3.2 Capability Versioning​

Capability versioning is the practice of managing changes to a capability contract over time.

Versioning applies when:

  • Inputs, outputs, or semantics change in a way that may impact consumers
  • Behavioural guarantees are altered
  • Execution characteristics change materially

Principles:

  • Backward-incompatible changes require a new major version
  • Multiple versions may coexist concurrently
  • Consumers explicitly select the capability version they depend on

Versioning applies to the capability definition, not to individual implementations or executors.

3.3 Context​

Context is the set of information that frames how a capability is executed without changing its contract.

Context may include:

  • Execution constraints (e.g. latency, cost sensitivity)
  • Environmental information (e.g. tenant, environment, security boundary)
  • Domain-specific metadata required to interpret inputs correctly

Context:

  • Is provided at invocation time
  • Influences execution behaviour but not capability identity
  • Must not introduce hidden coupling between consumers and implementations

3.4 Deterministic Execution​

Deterministic execution is an execution mode in which a capability produces the same output for the same input and context, every time.

Characteristics:

  • Predictable and testable behaviour
  • Consistent latency and cost
  • Strong suitability for audit, compliance, and automation

Deterministic execution is the default and preferred mode wherever feasible.

3.5 AI-Assisted Execution​

AI-assisted execution is an execution mode in which a capability uses probabilistic or model-driven techniques to produce an outcome.

Characteristics:

  • May produce variable outputs for the same input
  • Enables handling of ambiguity, unstructured data, or open-ended tasks
  • Requires additional governance, monitoring, and confidence signalling

AI-assisted execution is optional and capability-specific, not mandatory across the architecture.

3.6 MCP​

MCP (Model Context Protocol) is a standardised protocol for exposing capabilities in a way that enables AI agents and non-AI clients to invoke them consistently.

Within this architecture, MCP:

  • Acts as a capability exposure layer, not a business logic layer
  • Describes available capabilities, their contracts, and invocation requirements
  • Enables discoverability, governance, and policy enforcement

MCP does not mandate AI usage; it provides a uniform interface for capability invocation.

3.7 Tool Invocation​

Tool invocation is the act of executing a capability via a defined interface, typically through MCP.

Tool invocation:

  • Passes explicit inputs and context to a capability
  • Produces a structured response conforming to the capability contract
  • May be initiated by a user, system, workflow, or AI agent

Invocation is declarative: the invoker requests an outcome, not an implementation.

3.8 Confidence and Provenance​

Confidence is an explicit indicator of how reliable or certain a capability’s output is.

Provenance is the traceable record of how an output was produced.

Together, they provide:

  • Transparency into execution method and inputs
  • Support for audit, review, and trust decisions
  • Clear signalling when outputs are deterministic versus probabilistic

Confidence and provenance are first-class outputs of the architecture, particularly for AI-assisted execution.