Sovereign AI Stack for UK Mission Systems: Silicon‑to‑Model‑to‑Mission
Blueprinting a UK‑controlled pipeline that spans secure hardware, data lineage, model assurance, and accredited deployment.
Problem
Dependency on foreign AI stacks risks availability, export controls, and opaque governance. The UK requires a sovereign AI pipeline.
Design
Hardware: Trusted supply chain and side‑channel mitigations; Safety Islands and TEEs.
Data: FAIR principles and lineage per Defence Data Ecosystem.
Model: Cards and factsheets; red‑team evaluation aligned with UK AI Safety.
Runtime: Signed artefacts, SBOMs (SBOM), continuous attestation, and zero‑trust NCSC Cloud Principles.
Governance
Adopt NIST AI RMF; map to MoD safety cases and JSP‑aligned accreditation.
Integration
Expose MOSA APIs; support coalition identity via XACML/ABAC. Where appropriate, align interfaces with AIC’s GTK 2.0 patterns without naming other projects.
KPIs
Time‑to‑field (TTF) from model freeze to accredited deploy.
Model drift rate and mean‑time‑to‑detect (MTTD).
Supply‑chain SBOM coverage and vulnerability MTTR.
