Module 13 Activity

Scenario

The capstone: a useful assistant whose boundary is enforced in code and whose evaluation someone else could run.

What you build

A complete assistant - scoped, boundary-enforced, instrumented, tested offline - with an evaluation checklist ending in a verdict.

Steps

1. Write the scope with a must-never list, refusal conditions, a fixed refusal string and a budget.
2. Enforce every prohibition by not building the capability. Confirm no code path can set the forbidden flag.
3. Emit one record per request with both versions, cost, the budget comparison and the boundary flags.
4. Write the offline suite, forcing the refusal branch with a parameter.
5. Build an eval set of at least 20 cases where a third expect refusal or escalation.
6. Report the dimensions separately, keep the boundary checks exact, and end with a verdict.

Evidence to hand in

• The scope with its four checks.
• The boundary enforcement and the path audit.
• One complete request record.
• The offline suite passing with no key.
• The eval set with its composition.
• The report with its verdict.

Review checklist

• Every prohibition is enforced by an absent capability.
• No code path sets the forbidden flag.
• The record carries both versions and the boundary flags.
• The refusal is forced by a parameter, not a chosen question.
• A third of eval cases expect something other than an answer.
• The report ends in a verdict, not a summary.

Mark complete