AIPS — Governance Obligation & Risk Control Framework (Normative)
AIPS governance is derived from the actual ontology and impact profile of an AI Product version. Governance MUST NOT depend on a single overloaded label such as INFERENCE, GENERATIVE, or AGENTIC.
The governing rule is:
Risk is computed from orthogonal evidence. Governance obligations are then derived from the resulting risk posture and the Product characteristics that created it.
AIPS therefore separates:
- AI Product ontology — APT, IC, AR, EM, IE and topology;
- impact evidence — domain sensitivity, decision finality, action authority, scale and reversibility;
- Effective Risk Tier — R0–R4;
- Governance Obligation Profile — the controls that MUST apply to the specific Product version.
PrimaryArchetype and secondaryCapabilities remain deprecated compatibility vocabulary only. They SHALL NOT be used as the normative basis for new governance computation.
1. Governance Driver Declaration
1.1 Required ontology drivers (AIPS-GOV-001)
Every governed AI Product version MUST provide current values for the applicable ontology dimensions:
- APT — AI Product Type;
- IC — Intelligence Capability;
- AR — Agency / Autonomy;
- EM — Embodiment;
- IE — Intelligence Effect;
- Topology Scope;
- AI Qualification currentness;
- dependency currentness, especially for Derived AI Products.
These dimensions are independent. No single dimension acts as a replacement for the others.
1.2 Governance stacking (AIPS-GOV-002)
The effective obligations of an AI Product version SHALL be computed additively:
GovernanceObligations =
RiskTierObligations
UNION OntologyDimensionObligations
UNION ImpactObligations
UNION TopologyObligations
UNION QualificationCurrentnessObligations
UNION DependencyCurrentnessObligations
No obligation introduced by one driver may be suppressed solely because another driver appears lower risk.
1.3 APT treatment (AIPS-GOV-003)
APT does not itself establish a universal minimum risk tier.
PrimaryandDeriveddescribe intelligence provenance.Derived, especially Composite-Derived Products, introduces lineage, dependency-currentness and change-propagation obligations.- A Primary Product may be R4.
- A Derived Product may be R1.
Risk follows actual authority, effects, context and impact, not provenance labels alone.
2. Risk Tier Model
2.1 Canonical tiers (AIPS-RISK-001)
AIPS uses:
- R0 — Minimal
- R1 — Limited
- R2 — Moderate / Material
- R3 — High
- R4 — Critical / Escalation-required
Every AI Product version MUST expose exactly one computed EffectiveRiskTier.
2.2 Evidence-backed risk computation (AIPS-RISK-002)
The Effective Risk Tier SHALL be computed from evidence-backed RiskDriverAssessment records.
Conceptually:
effectiveRiskTier = MAX(
declaredRiskTier,
calculatedImpactTier,
agencyMinimumTier,
effectMinimumTier,
embodimentMinimumTier,
topologyMinimumTier,
domainSensitivityTier,
decisionFinalityTier,
actionAuthorityTier,
impactScaleTier,
reversibilityTier,
qualificationCurrentnessTier,
dependencyTopologyTier
)
Not every driver needs to contribute a distinct floor in every Product. The governance engine MUST retain the evidence and rationale for the drivers that materially contributed to the result.
A manually declared risk tier is an input, not authoritative final classification.
2.3 Risk Driver Assessment (AIPS-RISK-003)
Each material driver assessment MUST identify:
- driver kind;
- contributed tier;
- rationale;
- evidence reference where applicable.
The current machine-readable vocabulary includes:
DeclaredRisk, Impact, Agency, Effect, Embodiment, Topology, DomainSensitivity, DecisionFinality, ActionAuthority, ImpactScale, Reversibility, QualificationCurrentness, and DependencyTopology.
3. Ontology-derived minimum constraints
These are hard floors, not a complete risk calculation.
3.1 Autonomous action floor (AIPS-HC-001)
If:
AR = AutonomousAction
then:
effectiveRiskTier >= R2
3.2 Autonomous physical action floor (AIPS-HC-002)
If:
AR = AutonomousAction
AND IE includes PhysicalAction
then:
effectiveRiskTier >= R3
The Governance Obligation Profile MUST include:
ActionAuthorization;SafetyCase;EmergencyStopSafeState.
3.3 Closed-loop physical autonomy floor (AIPS-HC-003)
If:
EM = ClosedLoopAutonomous
AND IE includes PhysicalAction
then:
effectiveRiskTier >= R3
Physical safe-state and emergency-intervention controls are mandatory.
3.4 Enterprise autonomous orchestration (AIPS-HC-004)
If:
AR = AutonomousAction
AND topologyScope = EnterpriseOrchestrated
then:
effectiveRiskTier = R4
and the Governance Obligation Profile MUST include:
ExecutiveEscalation;CrossDomainCoordinationReview.
3.5 Action authorization (AIPS-HC-005)
If:
AR in {ActWithApproval, AutonomousAction}
then ActionAuthorization MUST be present in the Governance Obligation Profile.
4. Topology Scope
Every AI Product version MUST declare one topology scope:
StandaloneDomainCoupledCrossDomainFederatedEnterpriseOrchestrated
Topology describes the scope across which the Product can coordinate, influence, or propagate effects. It is a risk/governance driver, not an APT.
4.1 Topology change rule (AIPS-TOPO-001)
Any change in topology scope MUST trigger:
- risk re-evaluation;
- Governance Obligation Profile recomputation;
- dependency and authority analysis;
- deployment re-validation before promotion.
4.2 Undeclared runtime federation (AIPS-TOPO-002)
If observed runtime behaviour exceeds declared topology scope:
- the deployment MUST be marked non-compliant;
- risk MUST be recomputed;
- governance escalation MUST occur;
- the deployment MAY be suspended according to applicable policy.
5. Governance Obligation Profile
Every AI Product version MUST have exactly one current GovernanceObligationProfile for the evaluated version.
The profile records:
- the Effective Risk Tier from which baseline obligations were derived;
- governance driver dimensions used;
- the resulting obligation set;
- computation timestamp and method;
- supporting evidence references where applicable.
5.1 Baseline tier obligations (AIPS-GOV-100)
AIPS 0.1 establishes the following machine-checkable minimums:
R0–R1
No additional universal control floor is imposed beyond baseline Product qualification, lineage, security and applicable policy.
R2+
MUST include:
ChangeQualityMonitoring.
This is intentionally broader than model drift detection. For a symbolic Product it may monitor rule/knowledge changes; for a Derived Product it may monitor upstream dependency changes; for a learned model it may include data/concept/model drift.
R3+
MUST include:
InterventionOverride;ImmutableAudit.
InterventionOverride means an appropriate governed mechanism to stop, suspend, supersede, reverse, isolate, or otherwise intervene in Product operation. It is not limited to a graphical human override button.
R4
MUST include:
ExecutiveEscalation.
Additional controls may be imposed by domain, regulation, safety, or organizational policy.
6. Capability-specific obligations without capability-as-type
Intelligence Capability may add obligations, but does not become Product Type or create a universal risk floor by itself.
Examples:
- Generation may trigger provenance, harmful-content, copyright, prompt-abuse or output-verification controls.
- Reasoning may require explanation/evidence and logical-consistency evaluation.
- Planning may trigger plan-boundary, tool-authority and execution-validity controls.
- Control may trigger timing, stability, fail-safe and actuation constraints.
- Perception may trigger environment-coverage, sensor-failure and confidence-calibration controls.
The applicable obligations are additive and contextual.
7. Agency and effect controls
Agency and Intelligence Effect MUST be evaluated together.
Examples:
| AR | IE | Typical governance concern |
|---|---|---|
| Inform | Score | quality, explanation, misuse |
| Recommend | Decision support | human reliance, evidence, contestability |
| ActWithApproval | DigitalAction | authorization, approval integrity, audit |
| AutonomousAction | DigitalAction | delegated authority, containment, rollback |
| AutonomousAction | PhysicalAction | safety case, emergency stop, safe state |
The same IC may therefore have substantially different governance obligations depending on AR and IE.
8. Derived AI Product governance
Derived AI Products MUST additionally govern intelligence dependency currentness.
For a Derived Product:
- upstream AI Product versions MUST be explicitly traceable;
- qualification paths MUST remain current;
- material upstream changes MUST trigger impact analysis;
- Composite derivation MUST preserve contributor provenance;
- downstream Product Family qualification MUST be re-evaluated when qualifying primitive paths are invalidated.
DependencyCurrentness SHOULD be included where downstream qualification or behaviour materially depends on versioned upstream intelligence.
9. AI Qualification and governance currentness
Governance is version-specific.
If AI Qualification becomes STALE, REVALIDATION_REQUIRED, or INVALID, governance handling MUST follow the applicable lifecycle policy.
An INVALID AI Qualification MUST NOT be masked by a previous Governance Obligation Profile. Governance metadata cannot preserve AI Product Family membership once qualification evidence is invalid.
10. Runtime consistency and continuous assurance
Declared ontology and governance characteristics MUST match observed runtime behaviour.
Continuous assurance SHOULD compare runtime evidence with:
- AR;
- EM;
- IE;
- topology scope;
- execution mode;
- dependency graph;
- qualification currentness;
- risk driver assumptions;
- Governance Obligation Profile.
Material mismatch MUST trigger risk and governance recomputation.
11. AIPDS compilation
AIPDS MUST operationalize the current Governance Obligation Profile.
Examples include:
- monitoring rules;
- action authorization;
- approval gates;
- safe-state behaviour;
- immutable audit sinks;
- emergency intervention;
- dependency-currentness monitoring;
- runtime safety controls;
- escalation workflows.
A deployment MUST NOT claim conformance if it cannot realize obligations that are mandatory for the referenced AI Product version.
12. Legacy compatibility
The following 0.1 terms may remain serialized for compatibility but are deprecated as normative governance inputs:
primaryArchetype;secondaryCapabilities;- archetype-specific risk floors based on
INFERENCE,GENERATIVE, orAGENTICas mutually exclusive labels.
New implementations MUST use the ontology-aligned governance model described in this page.
13. Maturity assessment (Informative)
AIPCH maturity characteristics may inform assurance depth and governance oversight, but maturity scoring SHALL NOT override mandatory obligations derived from ontology, impact evidence, Effective Risk Tier, or applicable policy.