Unlabeled dismissed marketing risks, most recently dismissed first.

A batch rather than a single row so the client can prefetch the next document and step backwards without a round trip. Already-labeled risks are excluded, so two people working the queue only collide on the row they are both holding — which the create route resolves with a 409.

exclude_risk_ids is what makes skipping work. A skip writes nothing, so the row stays eligible and every later request would return it again — a labeler who skipped a whole batch would cycle those rows forever and never reach the rest of the backlog. The caller therefore has to say what it has already passed over. Bounded by MAX_EXCLUDED_RISK_IDS because it travels in the URL; past that the client drops its oldest entries, so a session with more skips than the cap may be offered its earliest ones again.