Glossary

Terms used across the AOE documentation, with the page that defines each one in full.

Product and packages

AOE — the product and engine family. Short for Agent Ontology Engine.

Agent Ontology Engine — the product category: a compiler and runtime for external domain models and corpora. See What AOE is.

Model Package — external declarations of a domain's schema and behaviour: types, fields, relations, projections, retrieval profiles, functions, actions, policies, validators, and migrations. See Package model.

Corpus Package — external units, assets, provenance, licence policy, and release configuration published against a model range.

Adapter Package — an external source or provider integration. It can add an importer, search provider, validator, evaluation provider, or action provider, and it cannot add a type or relation to the model.

Domain Package — the deployable composition of model, corpus, adapters, optional tools, and optional Skills. This is the boundary a reader or Agent installs.

Registry — the discovery and distribution service for Model, Corpus, Adapter, Domain, and Plugin Packages. The public site is a static discovery page rather than a Registry API.

Content and compilation

Unit — the domain-neutral source and IR item. A Model Package may give a Unit a more specific type name; Core does not need to know that name in advance.

Projection — a model-defined view of a Unit. Levels such as summary, core, and full each carry a token target.

Snapshot — an immutable compiled corpus release with verified identity. See Compilation and snapshots.

Snapshot identity — the binding of tenant, corpus, release, and content digest that the runtime recomputes before serving.

model.lock — the generated record of the exact resolved model version and schema digest a corpus was compiled against. Never edit it by hand.

Manifest — the generated corpus.manifest.json describing a release and its digests. Also generated, also never hand-edited.

IR — the stable intermediate representation the compiler emits. Core fixes the IR and the declaration meta-schema; everything domain-specific stays in packages.

Query and execution

Retrieval Profile — the model-declared configuration of candidate generators, features, constraints, and rerankers that a query is evaluated against.

Selection Plan — the query result. It records selected Units, score contributions, constraint decisions, relation expansion or exclusion, projection loads, budget use, and diagnostics. See Selection and execution.

Action — a model-declared, policy-gated operation. It defines input, output, capabilities, preconditions, policy, approval mode, idempotency, and evidence.

Execution Plan — the preflighted effects and required approvals returned before an Action runs. Also called an Effect Plan.

Principal — the identity on whose behalf a request is made.

Capability — the permission required to request a specific Action.

Idempotency — the guarantee that a repeated Action call with the same key does not repeat its effect.

Event Store — the append-only record of Action runs and their evidence, which keeps in-flight, awaiting-approval, completed, failed, timed-out, and replayed states distinguishable.

Evidence — the retained proof of what an Action did and what authorized it.

Agent-facing

Skill — an optional Agent-facing workflow or instruction layer. A Skill can teach an Agent when to request an Action. It is not a schema, a package boundary, or an authorization mechanism, and it cannot bypass the Action runtime.

MCP — Model Context Protocol, one of the three transports. See MCP transport.

Transport — how a client reaches the runtime: embedded SDK, MCP, or HTTP. All three share the same query, plan, and action contracts. See HTTP and Registry.

Distinctions worth keeping

A Unit is the neutral runtime item, while a type name is a Model Package concern. A Selection Plan can recommend an Action but cannot grant the capability to execute it. A Domain Package is the product boundary; engine packages are its implementation layers.

Avoid presenting one example domain's type counts, retrieval axes, or tools as universal behaviour. State the owning package when a detail is domain-specific. Writers should also read the terminology guide before adding new terms here.