AI Profile Execution

Status: Runtime-Managed Machine Configuration

AIProfile is a desktop-portable configuration package. Runtime may project an admitted profile into machine-local assets and descriptors, but an App does not bind an execution request to that profile. App requests carry caller identity, scenario content, and supported operation parameters; Runtime chooses the implementation.

Ownership Boundary

Responsibility	Owner
Portable AIProfile schema and validation	Desktop kernel
Profile installation and machine-local projection	Desktop / Runtime administration
Local asset and device-resource management	Runtime
Owner capability intent (AIConfig)	Exact App or Agent owner
Implementation selection and scheduling	Runtime
Execution diagnostics and audit evidence	Runtime

A LocalProfileDescriptor is machine configuration. It is not an App-facing model, route, connector, target, fallback policy, or reusable execution binding. The SDK does not translate an App's capability intent into one of these request controls.

App Call Path

1. The exact owner stores Local or Cloud capability intent in AIConfig.
2. The App invokes an admitted capability with identity, scenario content, and supported parameters.
3. Runtime evaluates its current machine configuration and selects an admitted implementation.
4. Runtime returns a typed result or typed failure. Any implementation details included in Runtime-issued diagnostics remain output evidence.

Changing an AIProfile can change how Runtime services later requests, but it does not rewrite owner intent and does not grant the caller implementation selection authority.

Machine Diagnostics

Runtime administration may report daemon reachability, installed-asset state, resource pressure, and profile installation failures. Internal resolution and scheduler probes can support those diagnostics. They do not become per-model or per-route readiness controls in an App, and an App does not run a preflight to choose an execution target.

Execution evidence is immutable audit information about what Runtime did. A caller must not feed a resolved model, route decision, connector, endpoint, or scheduling judgement back into another request as authority.

Scope Identity

AIScopeRef identifies the owner and surface for scoped configuration or records. It does not identify an execution implementation. A scope can retain Runtime-issued evidence without turning that evidence into a profile binding or request target.

See AI Scope Identity for the identity contract and AIConfig Surface for owner capability intent.

Reader Scenario: Install a Local Profile

1. A machine administrator validates and installs an admitted portable profile.
2. Runtime resolves the required local assets and records installation status.
3. An App keeps only its Local capability intent; it does not receive a model or route selector.
4. On the next capability request, Runtime decides whether the installed profile can supply an admitted implementation and either executes or fails closed.

Reader Scenario: Review Execution Evidence

1. An App submits a capability request without implementation controls.
2. Runtime chooses and executes an implementation.
3. Runtime emits typed diagnostics or audit evidence for the completed attempt.
4. The owner may display or retain that evidence, but cannot reuse it to pin a later request.

Key Points

• AIProfile and LocalProfileDescriptor are machine-configuration concepts.
• AIConfig is owner-scoped capability intent, not an execution profile.
• App requests contain no provider, model, route, connector, endpoint, target, fallback policy, profile descriptor, or readiness probe.
• Runtime alone owns implementation selection, resource scheduling, and execution evidence.
• Runtime-wide reachability does not imply readiness for a particular model or route.

Source Basis

• .nimi/spec/runtime/ai-provider.authority.yaml
• .nimi/spec/platform/core-protocol.authority.yaml