Your CI pipeline uses Claude to moderate user-submitted changelogs and release notes for policy violations before publication. After three weeks in production, your moderation logs show a recurring false-negative pattern: comments containing specific types of financial forward-looking statements pass the Specialized Advice category check even though your policy clearly prohibits them. The current implementation passes a flat list of category names to the moderation function. What change would most effectively eliminate this class of false negatives?