A layered architecture for AI-enabled defence and critical systems.
The reference model connects identity, provenance, compute, intelligent software, assurance and human decision rights so that critical authority and dependencies can be observed across the system rather than inferred from a model in isolation.
Identity & authority
Explicit identities and bounded permissions for people, devices, services and AI agents.
Data & provenance
Controlled data paths, lineage, sensor provenance, handling requirements and evidence.
Compute & infrastructure
Workload placement, administrative control, edge/cloud dependencies, recovery and continuity.
AI & autonomy
Models, agents and control functions operating within defined technical and policy boundaries.
Assurance & intervention
Evaluation, telemetry, change control, escalation and accountable human acceptance.

Operational control depends on seeing the dependencies beneath the interface.
Technical control surfaces
- Compute location and administrative authority
- Software, model and firmware supply chains
- Data residency, transfer and retention
- Keys, credentials and privileged operations
- Update, rollback and recovery mechanisms
Evidence surfaces
- System inventory and configuration state
- Evaluation and test results
- Runtime telemetry and exception handling
- Human intervention and escalation records
- Supplier and integration evidence
Control should be observable.
Identity, configuration, data provenance, software change, permissions and intervention decisions should produce evidence that can support technical review and institutional accountability.
Modular, reviewable and recoverable by design.
Explicit boundaries
Trust zones, delegated authority and integration boundaries are documented.
Least privilege
People, services and agents receive the minimum authority required for defined functions.
Evidence by design
Testing, logs, provenance and change records are treated as engineering outputs.
Recoverability
Continuity, rollback and intervention are designed before failure rather than after it.

