Skip to main content

Glossary

Use these terms precisely. In particular, proposal/authority, prediction/observation, and state/weights must not be used as synonyms.

A​

Autonomy binding​

AutonomyBindingV1: immutable mode epoch, enrollment revision, evaluation, context, and artifact fingerprints carried by proposal, request, policy, and authorization. It preserves provenance but is not authority by itself.

Autonomy enrollment​

AutonomyEnrollmentV2: exact owner-created scope for a v3 strategy, templates, privileges, targets, entities, capabilities, context/privacy, cognition route, risk, limits, budget, expiry, and fingerprints. Wildcards and overlapping resources are forbidden in the first release.

Autonomy mode​

Global AutonomyModeV1: observe, supervised, delegated, or paused. Mode grants no scope without an enabled enrollment.

Autonomy strategy​

Proposal-only v3 plugin role evaluated by Heart from AutonomyContextV1. It cannot authorize, dispatch, start a loop, or own executor/model/registry handles.

ACHIEVE​

Goal mode in which a goal becomes completed after its fulfilment has been demonstrated independently. Unlike MAINTAIN, the goal does not remain active as a continuous monitoring loop afterward.

ActionRequestV1​

Exact, typed request for one target and entity. It contains the capability, parameters, snapshot/world/target revision, preconditions, deadline, and an optional idempotency key. The request is not permission.

Adapter​

Plugin component that translates canonical Engine contracts to a target protocol and back. The public roles are separated more finely into provider, controller, executor, and oracle.

AuthorizationV1​

Scoped, expiring proof that one exact request type may be executed within a concrete scope. It binds, among other things, the request hash, target, entity, capability, limits, snapshot, mandate, and policy decision.

AutonomyProfileV1​

Legacy Homey-specific low-risk routine-delegation record. It remains typed audit and compatibility data but does not grant generic v3 authority. New generic scope uses AutonomyEnrollmentV2; yolo creates no profile.

B​

Behavior signal​

BehaviorSignalV1: plugin evidence about an external change or pattern, with scope, preference, context, provenance, and evidence grade. Repeated behavior is not consent.

Brain​

Replaceable deliberative component. The executive chooses a decision kind or semantic proposal; a specialist advises within capability families. A brain does not own authoritative world state, policy, authorization, an executor, or a success oracle.

Bounded context projection​

Temporary, bounded selection of a goal, current relevant world data, effect results, capabilities, and specialist metadata for a brain. The projection is not operational state and can be reconstructed.

C​

Capability​

A typed operation offered by a target under explicit input and effect schemas, units, preconditions, risk/privacy, limits, deadline, invocation mode, and recovery semantics.

Capability family​

Stable semantic family that groups dynamic target instances. The family is declared statically; unknown dynamic families remain opaque and read-only.

Capability graph​

Conceptual current collection of targets, capabilities, dependencies, and availability. The current runtime projects manifests and per-target discovery into snapshots/context; not every graph operation is exposed as a separate public CLI command.

Controller​

DomainController: translates a semantic proposal into an exact ActionRequestV1 within the capability envelope. This is not a realtime device controller in the sense of a motor or flight-control loop.

Coverage​

Description of the part of the world that an observation or provider output actually covers. Without complete relevant coverage, absence must not be interpreted as false.

D​

DEFER​

Policy outcome: the available evidence is currently insufficient, conflicting, or unsafe, so policy cannot return allow or a definitive deny.

Desired effect​

Typed desired state within a GoalSpec, bound to a capability family, entity selector, condition, and semantic parameters.

Deterministic executive​

Provider-free baseline that uses the same untrusted BrainDecisionV2 seam as a model executive. This allows known violations to be routed without an LLM.

E​

Effect oracle​

Plugin role that reconciles the proposal, pre-state, receipt, and fresh post-state. The oracle produces an EffectDeltaV1; it must return unknown when its measurements are insufficient.

EffectDeltaV1​

Durable description of observed change between two snapshots, including the evidence grade, achieved as true/false/null, observation IDs, and reason.

Entity​

Stably identified object in a target world, such as a warehouse bin or device zone. A name or embedding is not a canonical identity.

Evidence grade​

Classification as OBSERVED, DERIVED, INFERRED, UNKNOWN, CONFLICTING, or STALE. It is separate from confidence and quality.

Executive brain​

The single runtime-wide brain that chooses the next cognitive step type in a Heart application. Current composition: deterministic or one model-backed executive.

Execution receipt​

ExecutionReceiptV2: an executor fact about requested/accepted/running/succeeded/partial/failed/cancelled/unknown. A receipt is not proof that a GoalSpec effect was achieved.

Experience​

Historical actions, outcomes, and behavior evidence. Experience is not the same layer as current state, model weights, or temporary context.

G​

GoalSpec​

GoalSpecV2: declarative desired result with scope, desired effects, preferences, constraints, budgets, stop conditions, mode, and mandate binding.

Goal candidate​

GoalCandidateV1: non-authoritative request to instantiate one statically declared goal template inside exact enrollment scope. It is not a free GoalSpec.

Goal template​

GoalTemplateSpecV1: versioned typed executable-goal schema owned by a v3 plugin. Its inert compiler cannot mint a mandate or authorization.

H​

Heart​

The local, durable runtime core. The Heart observes, reconstructs state, evaluates goals and routines, calls brains when needed, validates proposals, runs through policy/authorization/execution, observes again, and stores audit data. The Heart is not an LLM or a hard-realtime controller.

I​

Idempotency key​

Key that allows a target or executor to recognize that the same logical request has been submitted again. This requires honest target semantics; the field alone does not make a non-idempotent physical action safe.

Dispatch attempt​

DispatchAttemptV1: durable pre-I/O record with a stable operation key, resource identity, lease fence, authorization expiry, and optional autonomy binding. An ambiguous prepared attempt is observed/reconciled after restart and is never blindly redispatched.

Imagined state​

Ephemeral counterfactual or predicted state. It may help planning, but it is not an observation and is not stored as authoritative world state.

Invocation mode​

immediate, task, or stream. Immediate returns directly; task has a durable handle plus polling/cancellation; stream assumes cursor/reconnect support. The current end-to-end reference covers task, not stream.

L​

Learning candidate​

Durable proposal to change a namespaced preference after validated behavior evidence. Candidate, shadow, promotion, and rollback are separate states.

Learning​

In Engine: bounded, auditable state/preference or routine adaptation. In the current runtime, it is not online training of model weights and cannot expand authority.

Lease​

SQLite-based exclusive ownership of the active runtime for one Engine store, with heartbeat and loss detection. It prevents two concurrent executive loops from operating on the same operational state.

M​

MAINTAIN​

Goal mode in which Engine continues to monitor a demonstrably desired state and may act again after observed drift.

Mandate​

StandingMandateV1: scope activated by an authorized actor for plugins, targets, entities, capabilities, limits, privacy, learning, validity, and manifest versions. A goal or brain does not create its own mandate.

Mini-brain​

A specialized learned component behind a capability contract. It requires, among other things, scope, uncertainty/defer behavior, provenance, a baseline, held-out evaluation, hardware measurements, and rollback. Being neural grants no special authority.

O​

Observation​

ObservationV1: typed evidence with entity, property, value, source, time, evidence grade, unit, quality, coverage, and optional artifact identity.

Opaque capability​

Dynamically discovered family that was not registered statically. It is projected as query-only, read-only, and observe-only until a typed manifest family is installed and enrolled.

P​

Plugin​

Installable engine.plugin/v3 package with a static manifest, explicit [autonomy], and runtime factory. V2 remains loadable without generic autonomy. A plugin owns domain semantics and adapters, not Engine's generic authority.

Policy decision​

Deterministic ALLOW, DENY, REQUIRE_APPROVAL, or DEFER with reasons. Only policy can cause an authorization to be created on the Heart path.

Prediction​

Predicted change. It remains separate from an independently observed effect.

Proposed action​

ProposedActionV1: untrusted candidate for one desired effect, bound to a goal and snapshot. It has no execution rights.

R​

Relation​

Typed, directed relation between two entities with source, time, and evidence grade. By contract, a RelationHypothesisV1 remains INFERRED.

Routine​

Durable activation layer above one linked GoalSpec. It contains a scoped guard, recurrence, cooldown, priority, conflict key, and status; it replaces neither policy nor the goal.

Routine compiler​

Plugin component that turns declared pattern semantics into inert RoutineSpec and GoalSpec data. It cannot create an authorization.

S​

Safe state​

Target-specific state to seek after failure. Reaching it must be observed again; Engine does not assume that rollback or stop succeeded.

Shadow​

Counterfactual evaluation phase without dispatch. A real opportunity exists only when the guard is true and the desired effect is demonstrably false; absence is not agreement. In autonomy OBSERVE, shadow is a real persisted strategy evaluation that structurally cannot dispatch. Neither form grants authority.

Suggestion​

SuggestionV1: free-form, non-operational model output. It cannot create a mandate, goal, proposal, request, authorization, or dispatch.

Specialist brain​

Plugin brain with declared supported families. It returns typed advice or a proposal, but no request authority or effect truth.

State​

Current target-bound facts and beliefs. Operational state is durable, typed, and reconstructible without a provider conversation.

T​

Target​

One concrete software or physical-system boundary under a stable target ID, such as one warehouse or home controller.

Target observation​

TargetObservationV2: one observation boundary for a target, with monotonic revision, entities, relations, observations, coverage, availability, and errors.

W​

Weights​

Reusable learned model parameters. They are not current world state, experience, or authority. The current bounded-learning path does not train weights.