Response-conditioned contrastive vector computation — supplies manufactured contrast sets to → Mean Difference (MD) vector extraction

Mean Difference's own recipe is deliberately minimal: collect two sets of contrasting example activations and average their difference, an approach that needs no optimization or gradient descent, only a legitimate two-group split to already exist. Persona-vector computation's real work happens before that arithmetic ever runs, in constructing a clean split that MD's generic recipe simply assumes as given. It generates 10 rollouts per extraction question under each of a contrastive instruction pair, then discards any rollout that fails to clearly express its intended condition — trait score above 50 for positive-prompted rollouts, below 50 for negative-prompted ones — before averaging (persona-vectors, §"2.2 Extracting persona vectors", p. 3). It also chooses, empirically rather than by default, where within each surviving rollout to read from: Appendix A.3 compares extracting from the final prompt token, an average over all prompt tokens, and an average over all response tokens, finding response-token averaging yields the strongest steering effect across evil, sycophancy, and hallucination, and adopting it as the pipeline's default (persona-vectors, §"A.3 Extracting persona vectors at different token positions", p. 27). Neither the filtering threshold nor the token-position choice is part of MD's generic definition; both are engineering decisions layered on top of it to manufacture the two-group split that MD's underlying arithmetic can then simply subtract and average.