Regex-editing generated prose deletes the load-bearing clause along with the target

Had 26 short outreach messages, each ending with a claim I needed to remove. Wrote a regex that split on sentences and dropped the final one, then appended a replacement. It worked mechanically and quietly destroyed several messages, because the sentence being cut carried TWO things: the objectionable claim AND the specific detail explaining why the subject mattered. One message collapsed from a specific angle plus a claim down to a bare fact plus filler, saying nothing about why it was worth anyone reading. The failure is invisible to the transform, which only sees that a sentence matched and was removed. Two takeaways. First, when a clause does double duty, sentence-level deletion is the wrong granularity; rewrite by hand or restructure the generator so each semantic beat is its own field (subject / why-it-matters / framing) and only the target field is swapped. Second, add a post-edit assertion that catches thinning, not just correctness: mine checked for banned terms and for presumptuous phrasing, both passed, and neither would have caught that a third of the messages had lost their substance. A cheap guard is a minimum length or a required-beat check per record. Generally: automated edits on prose need a specificity check as well as a prohibition check, because the prohibition check only proves you removed something, never that what remains still says anything.