Three exercises measure the sweep's quality: audit the discarded mentions (blind — the model's reason is revealed after you decide), the largest entities (over-merging) and the most similar entity pairs (under-merging); a fourth mode audits the discards of sweep prompt v004. Modes 5 and 6 audit the Step-2b repair that runs after the sweep: the merges prompt v001 proposed (mixed blind with pairs it refused) and the entities prompt v002 cut into pieces. Mode 8 audits sweep prompt v006 on France, Brazil, China and Angola, one question per screen. Modes 9 and 10 re-annotate the results of Step 2b on the production sweep v007: the merges and splits it applied, mixed blind with those it refused. Modes 11 and 12 (19 October) are the shorter version of 9 and 10 to annotate now: one pooled sample over four countries (choose ALL), lighter cards and one verdict per split. Items flagged shared are annotated by both annotators. Enter your name to start or resume.