Knowledge check and completion criteria

How to use this assessment

Complete the questions without referring to the chapter text. Then review the guidance and update your workshop artifacts where your answer exposes a gap.

This public knowledge check supports learning. It is not a professional certification examination.

Scenario questions

1. Local model claim

A team says its system is sovereign because an open-weight model runs on an on-premises GPU. Name four additional areas that must be assessed before that claim is useful.

2. Residency versus control

All application data is stored in the required country, but a remote provider can administer the service and controls the encryption keys. Which properties are established, and which remain unresolved?

3. System boundary

A RAG diagram shows the application, vector database, and model. It excludes the document source, identity provider, package repository, operators, backups, and logs. Explain why the boundary is incomplete.

4. Control quality

Rewrite “The model must be secure and compliant” as a testable control objective for approving a model version before deployment.

5. Evidence

A platform logs that a deployment succeeded, but the event does not contain the model hash, container digest, policy version, environment identity, or approver. Which evidence qualities are missing?

6. Air gap

Explain why an air-gapped runtime still requires an artifact supply chain and a controlled update process.

7. External dependency

Is the use of an external managed service automatically incompatible with sovereign AI? State the conditions under which it may be an accepted dependency.

8. Privacy-safe evidence

Design an audit event for a blocked RAG request without storing the complete prompt or retrieved document text.

Answer guidance

1. Strong answers cover data, model lifecycle, artifact dependencies, identity and egress, operations and recovery, evidence, ownership, and portability.
2. Residency may be established for the stated data. Administrative authority, key control, support access, dependency, and possibly operational autonomy remain unresolved.
3. The omitted components can change access, behavior, artifacts, recovery, or evidence and therefore materially affect control.
4. A strong objective identifies scope, approving authority, required provenance/license/evaluation/artifact checks, promotion mechanism, and retained evidence.
5. Attribution, version identification, completeness, and reproducibility are weak. The record may not support the claim that the approved release ran.
6. Disconnected systems still receive models, images, packages, drivers, policies, and fixes. The import, verification, promotion, installation, rollback, and evidence path must be controlled.
7. No. The organization must understand decision rights, access, data movement, continuity, portability, exit, evidence, and residual risk, and formally accept the dependency.
8. Record a request/event ID, timestamp, authenticated actor reference, system/model/policy versions, classification or rule triggered, decision, target component, and retention class. Include sensitive content only when specifically justified.

Artifact completion rubric

Mark each item complete only when another technically capable reviewer can understand it:

• The intended purpose and prohibited uses are unambiguous.
• Human and machine actors have named responsibilities.
• Material data, artifact, identity, administrative, and evidence flows are visible.
• External dependencies and decision rights are explicit.
• Control objectives are scoped, owned, and testable.
• Positive and negative control tests are identified.
• Evidence is attributable and versioned without unnecessary sensitive content.
• Gaps have owners, priorities, and completion evidence.

Completion outcome

SAI-100 is complete when the learner can explain why sovereignty is multidimensional and can produce the foundation package from the Foundation architecture workshop.