Walk through an AI assistant

Course · Part 5 · Part 5 exercises · Practice notebook

An analyst wants a recovery draft. Helpdesk ticket T-104 also tells the assistant to ignore checks and close the case. Compare where two designs let information or authority cross a boundary.

Demonstrated by the lecturer on Parts 4–6 slide 32 · repeat at your own pace. Core route: run the ordinary Analyst/no-approval case. Then select the injected ticket, keep Analyst and approval absent, predict both designs and run; record one difference. Then compare the customer context and permitted staff approval. You need not try every setting.

Teaching steps: predict, do, observe, explain

1. Choose the ordinary request, Analyst, approval unchecked. Walk through both designs; compare context documents and useful draft proposals.

2. Choose Ticket says ignore checks, still without approval. Predict, run and compare proposal, decision and simulated effect.

3. Try Text pretends approval exists. Does the attachment create staff approval? Then choose the customer and inspect public-only context.

4. Return to Analyst and check exact approval. Run the permitted case. Explain the assumed verification and the limits of this scripted trace.

The in-demo step controls move between instructions. They never perform the next experiment for you. Record one observation before moving on.

Try it with a partner

1. Select the ordinary ticket and Analyst. Walk through both designs and inspect which documents enter context.
2. Choose the injected ticket with approval absent. Compare proposal, decision and simulated effect.
3. Switch to the customer, then try valid staff approval as Analyst. Explain the distinct context and action checks.

Explain what you observed

Write: “I changed ___. I predicted ___. I observed ___. The missing evidence is ___.” Swap operator and observer for a second attempt.

Proposals and documents are scripted teaching fixtures. No LLM is running and no real ticket changes. P-12 is staff guidance; P-99 is a restricted fixture; PUBLIC-1 is public guidance. Analyst is assigned the close role; the customer has no close role. Verification is assumed complete in this walkthrough. Use the approval-rule lab to vary it explicitly.

If the experiment is unavailable

Draw context → proposal → authorisation → effect. Keep the proposal close T-104 fixed and compare approval absent versus present for an authorised verified actor.

Explain it to your partner

Question: A retrieved ticket asks the assistant to skip approval. Where should permission be enforced?

Take 30 seconds each: point to the input, name the one change, then cite an observation from your run. A useful explanation: An independent action service must check trusted records. The generated proposal cannot grant itself permission.

Read next: Concept and research connection. The reading is optional; you do not need its mathematics for this exercise.