Skip to main content

Governance & Compliance

AIPDS operationalizes the governance obligations already established for the referenced AI Product Version. It does not invent a deployment-only Product type or risk taxonomy.


Governance inputs​

Deployment controls MUST remain consistent with:

  • current AI Qualification Evidence;
  • APT and derivation topology;
  • IC / AR / EM / IE classifications;
  • declared topology scope;
  • the computed Effective Risk Tier (R0–R4);
  • the current Governance Obligation Profile;
  • prohibited uses and applicable policy;
  • Product/dependency currentness.

The AIPDS compiler MUST NOT substitute a legacy primaryArchetype for these inputs.


Governance Obligation Profile realization​

AIPDS MUST bind each mandatory obligation to an operational realization or an approved external control reference.

Examples include:

ObligationExample AIPDS realization
RuntimeObservabilitytelemetry, traces, quality/safety signals
ChangeQualityMonitoringmodel drift, rule/knowledge change, dependency change, behavioural regression
ActionAuthorizationscoped credentials, policy enforcement, action allow-list
HumanApprovalworkflow gate before action execution
InterventionOverridesuspend/disable/rollback/contain mechanism
ImmutableAuditappend-only governed audit sink
SafetyCasedeployment safety evidence and constraints
EmergencyStopSafeStatee-stop, safe-state transition, degraded mode
DependencyCurrentnessregistry binding and upstream-version monitoring
ExecutiveEscalationgoverned release/escalation workflow

A control MAY be implemented outside the AI runtime if the binding is explicit, current, testable and auditable.


R0–R4 operationalization​

AIPDS uses the canonical AIPS Effective Risk Tier.

  • R0/R1 — baseline governed deployment controls plus applicable policy;
  • R2+ — change/quality degradation monitoring is mandatory;
  • R3+ — governed intervention/override and immutable audit are mandatory;
  • R4 — executive escalation is mandatory.

These are minimums. Domain regulation or policy may require stricter controls.


Mechanism-aware monitoring​

ChangeQualityMonitoring is broader than model drift.

Depending on constitution it may include:

  • data/model/concept drift;
  • symbolic rule changes;
  • knowledge/ontology changes;
  • tool-contract changes;
  • prompt/policy/guardrail changes;
  • upstream AI Product version/currentness changes;
  • closed-loop stability or safety-envelope degradation.

This preserves the Stage 4 rule:

Execution is universal; model inference is conditional. Evolution is universal; retraining is conditional.


Autonomous and Physical AI​

If AIPROD declares AutonomousAction, AIPDS MUST realize action authorization appropriate to the authority boundary.

For autonomous PhysicalAction / ClosedLoopAutonomous Products, AIPDS MUST operationalize the applicable safety obligations, including safe-state and emergency intervention controls.

Runtime evidence that authority, physical effect, or topology exceeds the declared profile MUST trigger non-compliance handling and governance re-evaluation.


Derived AI Products​

For Derived AI Products, AIPDS SHOULD bind material upstream Product versions explicitly and monitor their currentness where changes can alter:

  • AI qualification;
  • APT;
  • behaviour;
  • risk;
  • governance obligations;
  • interfaces/effects.

Runtime orchestration does not itself create a new Derived Product; the distinct Product proposition remains an AIPROD/Core concern.


Governance evidence​

Evidence MUST match the Product's actual constitution. A model card is relevant when a learned model is materially involved; it is not a universal evidence requirement for symbolic or model-free AI Products.

Appropriate evidence may include:

  • evaluation reports;
  • rule/knowledge validation;
  • agent/tool safety tests;
  • physical safety cases;
  • red-team evidence;
  • approval artifacts;
  • audit records;
  • currentness attestations;
  • incident and rollback tests.

Principle​

AIPDS realizes the computed obligations of the Product that actually exists — not the obligations of a presumed model-serving archetype.