Changing a prompt can fix one behavior while breaking another. Compare the old and new instructions on the same frozen cases, then record both new failures and repaired failures. Replacing the baseline with the new result erases the evidence you need to judge the change. In this lesson you specify a regression comparison and preserve the baseline until its owner approves replacement. A safe comparison also records the fixture and instruction versions. If the tool or model changes, name that change rather than attributing every difference to the prompt. The exercise uses a local deterministic session so the effect of an instruction boundary stays visible.
A worked artifact for a separate fictional task. Compare its actions with the contract.