Choose 200 of a Million Records to Label, Improving on Uniform Random Sampling

You are given a fixed collection of roughly $$10^6$$ unlabeled records. You may request labels for exactly $$200$$ distinct records, and the entire batch must be chosen at once. You cannot label a small subset first, inspect those labels, and then choose the rest of the batch. The requester asks for the “useful” records but offers no definition of useful. Selecting $$200$$ records uniformly at random would be accepted as a fallback, but it is considered a weak answer.…