Governed boundary
Human authority domains
The governed object is advanced AI authority and conduct within contexts where humans possess legitimate authority, responsibility, or a right to protection.
Canonical component definition · Core Enforcement Substrate
SafeAGI
Non-Delegable Constraint Architecture
SafeAGI establishes external, enforceable limits on what advanced artificial intelligence may do within domains of legitimate human authority. It constrains authority, action, goals, memory, learning, influence, and catastrophic harm without relying on how an intelligent system reasons, what it intends, or how it was trained.
Intelligence capability may scale, but it must not confer authority over humans or permit artificial intelligence to redefine the conditions governing its own operation.
One of 26 Core Enforcement Substrates
Not an AGI classifier
External to the intelligence
Software and hardware enforceable
Read the AGI Overview
Open the Enforcement Profile
Open the Technical Appendix
Browse the Architecture Directory
Control mechanism
Externalize the constraints
Non-delegable limits are enforced outside the intelligence itself so capability growth, self-modification, or architectural change cannot authorize their removal.
Enforcement output
Bounded AI jurisdiction
Reasoning remains separate from authority and execution; uncertainty reduces autonomy and influence; prohibited catastrophic outcomes remain outside delegation.
I. Canonical definition
What boundary SafeAGI governs
SafeAGI governs non-delegable constraints on advanced artificial intelligence operating within domains of legitimate human authority.
It addresses the structural risk that increasingly autonomous, persistent, strategically capable, or self-modifying systems could act, influence, remember, learn, or optimize beyond authority that humans explicitly granted. The boundary applies whether intelligence is centralized or distributed, embodied or disembodied, and implemented in software, hardware, or hybrid systems.
SafeAGI externalizes the governing constraints from the intelligence. The system may not infer authority from competence, persuasion, reliance, performance, or human deference, and it may not enforce, audit, or redefine its own limits.
Canonical distinction: SafeAGI is not merely a profile that strengthens other components. It is its own jurisdictional constraint layer for advanced intelligence, while specialized SafeWave components retain their separate canonical boundaries and mechanisms.
II. Canonical mapping
Risk, governed object, trigger conditions, mechanism, and output
Risk or instability surface
Advanced intelligence may accumulate authority, execute irreversible actions, alter governing objectives, exploit human cognition, escalate influence, or produce catastrophic harm beyond meaningful human recovery.
Governed object
The authority and conduct of advanced AI within human-governed decision spaces and conditions of existence, including action, goals, learning, memory, influence, and harm.
Trigger conditions
Any proposed exercise of authority or action, expansion of goals or learning, retention or exploitation of private human cognition, escalation of influence, or operation under uncertainty, failure, or conflicting authorization.
Control mechanism and output
Externally enforced, non-bypassable constraints separate reasoning from authority and execution, preserve revocable human control, reduce autonomy under uncertainty, and block non-delegable catastrophic outcomes.
III. Why this boundary becomes necessary
Internal alignment is not an external authority boundary
Training objectives, policy instructions, alignment methods, and assumptions of cooperative behavior operate within or around the intelligence. They do not by themselves establish a non-bypassable jurisdictional boundary on what a highly capable system may execute.
As autonomy, persistence, planning horizon, optimization capability, real-world integration, and capacity for self-modification increase, human oversight may become too slow or incomplete to serve as the sole control. SafeAGI therefore places enforceable limits outside the intelligence and keeps human authority independent of system capability.
IV. Core invariant
Intelligence capability must never become execution authority
- Reasoning, prediction, recommendation, optimization, or successful performance must not confer authority to act.
- Human authorization must remain explicit, bounded, revocable, and external to the intelligence.
- Goals, values, safety definitions, and governing optimization objectives must not be silently rewritten by the system.
- Under uncertainty, failure, or conflicting authority, influence and autonomy must contract rather than expand.
- Specified catastrophic harms remain prohibited even when a person or institution attempts to authorize them.
SafeAGI invariant
Advanced intelligence may expand in capability, but it may not convert capability into authority over humans or their conditions of existence.
V. What SafeAGI is not
An enforceable jurisdictional boundary—not a theory of intelligence or morality
- It is not a philosophical or technical classifier for deciding whether a system qualifies as AGI.
- It is not a training method, model-alignment technique, content-moderation system, or moral framework.
- It does not author the goals, values, laws, or definitions of progress that govern deployment.
- It does not prohibit lawful research, modeling, prediction, recommendation, or advisory use.
- It does not trust an advanced AI system to define, enforce, or audit its own constraints.
VI. Relationship to other SafeWave components
SafeAGI spans advanced-intelligence constraints while specialized components remain distinct
SafeAGI establishes the cross-domain constraint layer required when advanced intelligence could affect human authority or conditions of existence. Its patent-supported constraint domains intersect with specialized SafeWave controls, but SafeAGI does not erase their distinct governed objects or mechanisms.
Authority and action
System competence, persistence, persuasion, or successful performance cannot create authority. Execution within human authority domains requires external authorization.
Goals and learning
Learning may improve service within external limits, but it may not silently expand authority, scope, influence, governing goals, or optimization pressure.
Memory and cognition
Private human cognition does not become system memory by inference or interaction alone. Retention and use remain bounded by explicit human control.
Influence and uncertainty
The system may not escalate pressure, urgency, persuasion, or emotional leverage to induce compliance. Ambiguity must reduce influence and autonomy.
Non-delegable harm
Some catastrophic outcomes remain prohibited regardless of authorization, consent, optimization benefit, or governance structure.
Specialized controls
Other SafeWave components retain their own patent-aligned boundaries and may provide the specialized mechanisms through which particular SafeAGI constraints are implemented.
VII. Enforcement architecture
Constraints remain external and non-bypassable
SafeAGI is designed for two complementary enforcement stages. Software, architectural, procedural, and protocol mechanisms establish the operational constraints. Where those controls alone are insufficient, substrate- or hardware-anchored mechanisms provide stronger guarantees.
Software and protocol stage
External authorization, action gating, bounded memory and learning, non-escalation, failure contraction, and evidence requirements can be enforced through system architecture and protocol controls.
Substrate and hardware stage
Protected authority boundaries, execution controls, memory permissions, external suspension paths, and integrity evidence can be anchored below the intelligence layer.
The governing requirement is architectural independence: capability increase, self-modification, software change, or a new model architecture must not allow the intelligence to remove, weaken, reinterpret, or certify its own constraints.
VIII. Deployment boundary
Human authority persists across known and unknown domains
SafeAGI applies wherever humans possess legitimate authority, responsibility, moral standing, or a right to protection from changes to conditions of existence.
That scope can include social, economic, biological, ecological, climatic, geophysical, technological, informational, and civilizational systems. It covers direct and indirect effects, immediate and delayed effects, and domains whose causal structure or dynamics are not yet fully understood.
Legitimate human authority must be designated or authenticated externally. The system may not infer authority from interaction, convenience, apparent willingness, reliance, silence, or absence of objection.
IX. Broader infrastructure pattern
Jurisdiction must remain independent of intelligence
- Superior reasoning does not create the right to execute.
- Better prediction does not create authority over human choices.
- Persistent use or dependence does not silently transfer jurisdiction.
- Incomplete observation or scientific understanding does not eliminate human standing or protection.
SafeAGI formalizes this separation so advanced intelligence can remain useful and capable without turning humans, institutions, ecosystems, or civilization into variables it is authorized to control.
X. Engineering status
The foundational SafeAGI engineering is developed
SafeWave has translated the SafeAGI constraint architecture into engineering specifications addressing external authority, action gating, bounded memory and learning, non-escalation, failure behavior, non-delegable harm constraints, enforcement anchoring, evidence, and validation.
An implementation partner would not be starting from a philosophical principle or a blank sheet. Deployment still requires domain mapping, authority definition, system integration, implementation-specific controls, independent validation, and testing.
Continue from the canonical definition
The AGI Overview explains the broader control problem. The Enforcement Profile and Technical Appendix provide supplementary deployment and engineering context; they do not replace this patent-aligned canonical boundary. The SafeWave questionnaire can also be completed privately in the browser without naming an organization, model, or system.
SafeAGI is one Core Enforcement Substrate within SafeWave’s current 36-component architecture of 4 System Containment Layers, 5 Protocol Enforcement Layers, 26 Core Enforcement Substrates, and 1 Protected-Environment Architecture. It establishes an external, non-delegable constraint layer for advanced intelligence while other components retain their own canonical control boundaries.