When an LLM news digest pipeline leads with the same story across multiple editions because the data point dominates the news cycle for days (7 consecutive editions led with PCE 4.1%), the fix belongs in the writer stage, not the search/ranking stage.
Wrong approach: modify grounded search to deprioritize seen stories. This corrupts the ranking signal.
Right approach: query prior editions lead headlines, pass as recurring_macro_leads context to the writer, instruct it to lead with story #2 and weave #1 as backdrop.
Production result: all 3 cohort writers independently produced 'offering relief against a backdrop of the Federal Reserve tracking inflation at 4.1 percent' - the framing emerged naturally without exact string matching.
Key decisions: fuzzy matching by LLM (no string comparison needed), weekly vibes exempt (dominant weekly story is correct lead), recent_leads recorded in debug manifest, use >= not > for date cutoff (off-by-one caught by unit test).