Module 05 Activity

Scenario

The crew is about to be trusted with something that cannot be undone. This is the layer that makes that defensible.

What you build

A guardrail that fails closed, assertions in callbacks, one well-placed human gate, and rejections an agent can act on.

Steps

1. Write a guardrail checking parse, shape and policy, with the policy limits as data at the top of the file.
2. Wrap it so that an error in the check BLOCKS rather than allows. Prove it by breaking the checker deliberately.
3. Attach assertions to task callbacks - including one that flags hedging language in a decision output.
4. Sort your steps by reversibility and place the minimum number of gates, immediately before the irreversible steps.
5. Design the rejection format with three fields: what was wrong, what is correct, what to do next. Use it for human reviewers too.
6. Measure what happens after a rejection - does the retry differ from the first attempt?

Evidence to hand in

• The guardrail with policy limits as data.
• Evidence that a broken checker blocks rather than allows.
• The callback assertions and what each catches.
• The reversibility table with the chosen gates.
• The three-field rejection form.
• Before-and-after outputs for one rejection.

Review checklist

• Policy limits are data, not prompt text.
• The guardrail fails closed, demonstrated by breaking it.
• One callback assertion catches hedging in a decision output.
• Gates sit immediately before irreversible steps, and the count is minimal.
• The retry after a rejection is measurably different from the first attempt.

Mark complete