Control AI outputs and actions at runtime
A realization may predict, classify or propose without automatically being allowed to act. AMF separates fabrication and evaluation from runtime checks: output schemas, permissions, refusal and escalation under mission-specific rules.
A private AI system combines models, data and tools.
Models & experts
Specialized learned components, large or small, retained only when the mission justifies them.
Data & knowledge
Sources, corpora, memory, graphs, documents and information assets with explicit provenance and status.
Rules & verifiers
Deterministic constraints, schema validation, exact computation, policies and independent controls.
Tools & execution
APIs, business systems and authorized actions inside explicit operational boundaries.
Routing & policy
Allocation of responsibilities, fallback, refusal, escalation and authorization according to context.
IQ Fabric runtime
Execution of frozen QIRs with receipts and without silent mutation of the qualified identity.
A model suggestion is not permission to act.
AI proposal
A model or composed realization emits a prediction, structured response or suggested action.
Runtime checks
Apply available schema, label and business-policy checks; a malformed or uncertain output can be refused or escalated.
Separate authorization
Allow an external action only when the executing actor, context and permissions authorize it; otherwise block or refer.
Trace & review
Record the decision and permitted evidence, subject to data handling rules; use observations to investigate failures.
A structured suggestion meets the configured schema and authorization checks. Only the permitted action may continue.
The AI may suggest an action without gaining permission to perform it. The execution is blocked or referred.