A support-summarization feature is in production. The team wants to rewrite the system prompt to improve tone, but has been burned by changes that fixed one case and broke others. The summaries are free-form text, so exact-match comparison is not possible. What is the best way to decide whether to ship the new prompt?