Selection and execution

Reading ontology-backed knowledge and changing state are separate contracts in AOE.

Selection Plan

A query request is evaluated against an external Retrieval Profile. Candidate generators may use lexical, facet, graph, or plugin-provided signals. Features, hard constraints, soft constraints, relation semantics, reranking, projection choice, load order, and token budget produce a Selection Plan.

The plan records selected Units, score contributions, constraint decisions, relation expansion or exclusion, projection loads, budget use, and diagnostics. Visibility is enforced before candidate providers receive data so a private Unit cannot leak through a relation or score.

Execution Plan

An Action starts from a model declaration and a Request Context. The runtime validates inputs, principal, capability, preconditions, provider binding, side-effect class, idempotency, and policy. Preflight returns an Effect Plan without executing the provider.

Execution proceeds only after required policy or human approval. The run and its evidence are appended to the Event Store so in-flight, awaiting-approval, completed, failed, timed-out, and replayed states remain distinguishable.

The boundary

A Selection Plan can recommend a Unit or Action. It cannot grant the capability to execute it. A Skill can teach an Agent when to request an Action. It cannot bypass the Action Runtime.