SAI-280 knowledge check

Scenario questions

1. Why is model output a proposal rather than authorization?
2. What belongs in a governed tool registry?
3. Which identities must remain distinguishable during a run?
4. When must action authorization be rechecked?
5. What makes human approval sufficiently specific?
6. Which boundaries apply to durable agent memory?
7. How should a run receipt represent partial success?
8. What should happen when policy or evidence services fail?

Completion criteria

The design passes when purpose, authority, identities, tools, typed inputs, permissions, approvals, confinement, memory, budgets, stop conditions, receipts, evaluation, failure, compensation, and incidents are explicit and enforced outside model instructions.

Completion confirms conceptual readiness only. Practical assessment remains unavailable while SAI-280 is in roadmap status.