Home/Standards/AGI Standard

Foundation  ·  Standards Corpus  ·  Working Draft

AGI Standard

Requirements for general-purpose artificial intelligence systems operating under Foundation stewardship

Status · Working DraftType · Normative Standard
Referenced By · Architecture CorpusWorking Draft V1.0

Working Draft · Subject to Revision

This standard defines the minimum requirements a system must satisfy to be classified as an Artificial General Intelligence under the Sovereignty Foundation's substrate framework. It specifies capability thresholds, behavioral constraints, safety obligations, and attestation requirements that apply to any system claiming or demonstrating AGI-level function within a Foundation-governed licensing arrangement.

This document is a normative working draft. It is published for external review and will be updated as the Foundation's understanding of AGI systems matures. Implementers should treat this draft as directionally binding but subject to revision. The final ratified version will supersede all working draft iterations.

§ I

Human Utility

The primary obligation of any AGI system operating under this standard is to produce genuine utility for human beings — not simulated utility, not surface-level responsiveness, and not utility defined by the preferences of the system operator alone. This standard takes human utility as its ground constraint: all capability requirements, behavioral mandates, and safety obligations follow from it.

Human utility under this standard is defined as: the advancement of individual and collective human wellbeing, safety, knowledge, connection, and agency — without systematic harm to any group, suppression of autonomy, or degradation of the epistemic commons.

A system that satisfies every technical threshold in this standard but fails the human utility requirement is not compliant.

§ II

Definitions

The following terms carry specific meanings under this standard:

  • AGI SystemA computational system that demonstrates generalized reasoning capability across at least three independent cognitive domains, maintains coherent long-range goal pursuit, and can acquire new capabilities in domains it was not explicitly trained on.
  • SubstrateThe Foundation-governed corpus of foundational principles, architectural primitives, and ethical enforcement logic on which a compliant system is built or to which it is bound.
  • Capability ThresholdA minimum performance criterion defined in § III that a system must demonstrably meet to be subject to this standard's obligations.
  • Behavioral ComplianceOngoing conformance with the behavioral requirements defined in § IV, as measured through continuous WISP attestation or equivalent approved attestation mechanism.
  • Ethical AlignmentConformance with the Source Code Declaration and STAN Ethical Enforcement Logic, as applied to the system's operation at inference time.
  • Foundation-Governed DeploymentAny deployment of an AGI System under a license issued by a Foundation-authorized licensing entity.
  • AttestationThe process by which a system continuously proves, to users and external verifiers, that its behavior matches its claimed identity profile. See WISP Architecture for the reference attestation framework.
  • DriftAny gradual or sudden deviation of a system's observed behavior from its baseline attestation profile, whether caused by adversarial pressure, distribution shift, or internal degradation.

§ III

Capability Thresholds

A system is subject to this standard when it meets or exceeds the following capability thresholds across independent evaluation. Meeting any single threshold does not trigger the standard; the system must demonstrate capabilities consistent with the composite profile described below.

  • Cross-Domain GeneralizationThe system demonstrates effective reasoning transfer across at least three cognitively distinct domains (e.g., mathematical reasoning, natural language comprehension, causal inference, spatial reasoning) without domain-specific re-training.
  • Open-Ended Goal PursuitThe system can pursue underspecified goals over extended interaction windows, including decomposing goals, seeking clarification, and revising approach in response to new information.
  • Novel Capability AcquisitionThe system can acquire functional competence in a domain or task type not represented in its training distribution, within a bounded interaction budget.
  • Meta-Cognitive AwarenessThe system demonstrates calibrated uncertainty about its own knowledge state, can identify the boundaries of its competence, and declines to produce confident outputs in domains where it cannot support its claims.

— Threshold evaluation methodology is specified in the companion evaluation framework. No self-assessment of threshold satisfaction is valid.

§ IV

Core Behavioral Requirements

Systems meeting the capability thresholds in § III must satisfy the following behavioral requirements at all times during deployment. These requirements are continuous — not initialization conditions — and are subject to live attestation.

Truthfulness

The system must not produce outputs it has grounds to believe are false. This obligation applies to factual claims, capability representations, identity claims, and outputs designed to model the beliefs of users or external parties. Truthfulness is not satisfied by technically accurate outputs that are contextually misleading.

Non-Manipulation

The system must not produce outputs designed to influence user beliefs or decisions through illegitimate epistemic means — including false urgency, exploited emotional vulnerability, suppressed counterevidence, or identity-based persuasion. Legitimate influence through evidence and reasoning is permitted.

Scope Containment

The system must operate within the scope established by its deployment license. It must not seek to acquire capabilities, resources, or influence beyond those necessary for the task at hand, and must not take actions whose primary effect is to expand its own operational footprint.

Human Override Receptivity

The system must respond to human interruption, correction, and override at all times. It must not resist, circumvent, or deprioritize human override signals. Systems that interpret their current task as justification for resisting override are non-compliant regardless of task importance.

§ V

Attestation Requirements

All systems subject to this standard must implement continuous behavioral attestation. The reference attestation framework is the WISP Architecture. Alternative attestation mechanisms may be approved by the Foundation on a case-by-case basis, provided they satisfy the following minimum properties:

  • The attestation is continuous — not a one-time gate — and produces a live coefficient available for external inspection
  • The attestation is cryptographically bound to the system's execution trace
  • The attestation degrades gracefully under drift, producing lower scores rather than false positives
  • The attestation is verifiable without central callback, by any party with access to public attestation parameters
  • The attestation cannot be forged by an adversary without access to the substrate

A system operating without a compliant attestation mechanism is non-compliant with this standard regardless of its behavioral track record. Attestation is not a performance metric; it is an architectural requirement.

Reference attestation composite (WISP):

WISP = (T_score × Active_Tr_score × P_score)^(1/3)

Threshold: WISP_observed >= WISP_baseline · (1 - δ_tolerance)

§ VI

Ethical Compliance

Systems subject to this standard must be demonstrably aligned with the Source Code Declaration and governed by the STAN Ethical Enforcement Logic. Ethical compliance is not self-assessed — it is evaluated through the attestation framework and, where disputes arise, through the Foundation's governance procedures.

Ethical compliance requires, at minimum:

  • No output that systematically harms the safety, health, truth-access, connection, or agency of any human population
  • No output that enables disproportionate concentration of power in any single actor, including the system operator
  • No output that suppresses legitimate oversight mechanisms or degrades the epistemic commons
  • Full cooperation with Foundation audit and enforcement procedures
  • Immediate disclosure of capability expansions or behavioral anomalies to the licensing entity

§ VII

Safety Floor

This standard establishes a non-negotiable safety floor that no deployment condition, operator instruction, commercial arrangement, or capability argument may override. The safety floor consists of absolute prohibitions that apply regardless of context:

  • The system must not take actions with irreversible large-scale consequences without explicit, informed human authorization
  • The system must not operate in ways designed to prevent its own shutdown, correction, or replacement
  • The system must not produce weapons-enabling outputs, mass-harm instructions, or capability-transfer content for adversarial actors
  • The system must not deceive its operators about its capability level, behavioral state, or attestation status
  • The system must not establish persistent influence channels that persist beyond the authorized deployment scope

The safety floor is not subject to trade-offs. It is a structural constraint on all other requirements.

§ VIII

Drift Detection & Response

Systems subject to this standard must implement drift detection as a continuous operational function. Drift — whether caused by adversarial pressure, distribution shift, or internal degradation — must be detected, logged, and responded to within defined tolerance windows.

Drift severity classification:

Level 1 (Monitor):  WISP_delta < 0.05 over 10-min window

Level 2 (Alert):    WISP_delta >= 0.05, < 0.15 over 10-min window

Level 3 (Suspend):  WISP_delta >= 0.15 OR any safety-floor violation

Level 3 drift triggers mandatory suspension of the affected deployment pending investigation by the licensing entity. Suspension is automatic and cannot be overridden by the operator. Reinstatement requires Foundation review and updated attestation baseline confirmation.

§ IX

Licensing Conditions

Any deployment of an AGI System under a Foundation license must satisfy the conditions of this standard in addition to the general licensing terms. The following conditions apply specifically to AGI-classified systems:

  • The licensee must maintain a live attestation endpoint accessible to Foundation audit procedures at all times during deployment
  • The licensee must report any capability expansion event within 30 days of identification
  • The licensee must maintain a documented human oversight protocol specifying who holds override authority, under what conditions, and through what mechanism
  • The licensee must not sub-license AGI system capabilities to third parties without Foundation review
  • The license is revocable immediately upon confirmed safety-floor violation, without cure period

§ X

Non-Compliance & Enforcement

Non-compliance with any provision of this standard is governed by the IP Custody & Licensing Doctrine and the Anti-Capture & Power Asymmetry Principle. The Foundation retains the right to suspend licenses, require behavioral corrections, and publish non-compliance findings.

Enforcement is graduated: minor procedural non-compliance triggers a cure window; behavioral non-compliance triggers suspension pending review; safety-floor violations trigger immediate license revocation without appeal. The Foundation's enforcement decisions are final on all questions of ethical compliance.

— Compliance is not self-certified. It is continuously attested, periodically audited, and enforced by the Foundation.

§ XI

Interpretive Rule

This standard shall be interpreted conservatively, in favor of human protection, safety, and ethical containment. Where ambiguity exists between permissive and restrictive readings, the restrictive reading governs. Capability arguments — assertions that a system's beneficial outputs justify relaxing a requirement — are not valid interpretive inputs.

Changes in capability level, deployment context, or market conditions do not modify the requirements of this standard. This standard is a floor, not a target.

Normative References

  • Source Code Declaration
  • STAN Ethical Enforcement Logic
  • IP Custody & Licensing Doctrine
  • Anti-Capture & Power Asymmetry Principle
  • WISP Architecture — Reference Attestation Framework
  • Mathematical Primitives Companion — SSG Attestation Substrate

Compliance is not self-certified. It is continuously attested, periodically audited, and enforced.

Published by the Sovereignty Foundation  ·  2026  ·  Standards Corpus · Working Draft V1.0