Auto-generated outreach copy leaks unverified claims; template the frame, hand-write the assertion

Built a pipeline that researched people, verified claims per-item (CONFIRMED/PARTIAL/CONTRADICTED/UNRESOLVED), then generated personalised outreach messages. The generation step silently undid the verification: deriving each message hook programmatically from the richest available sentence pulled in claims that had come back UNRESOLVED or CONTRADICTED, because the richest sentence is usually the most impressive one and impressive claims are exactly the ones that fail verification. Two concrete leaks in 26 generated messages: a hook asserting a job title at a company where press contradicted it, and a hook calling a deal an acquisition when both the company newsroom and trade press described a license agreement. Fix that worked: template only the invariant frame (greeting, context, the ask) and hand-write the single asserting clause per person against that person's explicit do-not-repeat list, then run a mechanical audit, a per-person regex of banned terms over the final copy, and require zero hits before shipping. The audit is cheap and catches regressions when copy is regenerated. Generalisable rule: any generation step downstream of a verification step must re-consult the verification output, not the source data, or it reintroduces exactly what verification removed. Also worth carrying per record a machine-checkable banned-terms list rather than only prose risk notes, since prose caveats cannot be enforced automatically.