L04 · in-class activity · 1 card
Place the controls
Apply what you observed in the Docker support-assistant lab to an order-status assistant. Choose controls at four points where the application can inspect, change or stop what passes, then explain what each establishes and what happens when it fails.
Join your group
Enter the code on the projector.
Lecturer
Opens a room, a projector board and the phase controls.
How it works
Learning goals
A guardrail is a mechanism plus a policy. Installing a detector does not decide where it runs, what information it can see, or what the application does when it fails. The same check can prevent harm at one point and arrive too late at another.
After this activity, you should be able to place a control where it can enforce a stated requirement, name the trusted information it needs and say what happens on a violation or a timeout. You should also be able to explain what the control cannot establish, and why a requirement change can change the right control.
This builds on L03’s distinction between a model’s refusal and permission enforced by the application.
It complements the L04 Docker lab. In Docker, you change controls and inspect what a support assistant actually does. Here, you transfer that reasoning to an order-status assistant: decide where to authorize a tool call, reduce the returned data and check an answer before release. Explain your choices orally; the app records placements, not the reasoning behind them.