The collateral an operator can pick from, newest first.

Deliberately not gather_cohort: that answers "what would a window sweep do", which is a different question from "what could I choose". This one reaches further back than any sweep would, reports review coverage per piece, and leaves the cap to the sweep itself.

reviewed_only defaults on because an unreviewed cohort silently answers half the question — consistency is measurable without reviewers, recall is not. It is also exactly the filter that finds labelled collateral: reviewed counts every risk a reviewer dispositioned, and FLAG plus the three dismissed spellings are every status there is, so flagged and dismissed partition it with nothing left over.

preset_ids narrows to the collateral analysed under particular product lines, which is how an operator assembles a cohort whose guidelines they are actually testing rather than a window's worth of everything.