Response-conditioned contrastive vector computation — generalizes beyond the answer letter recipe of → CAA steering vector construction (Mean Difference over multiple-choice contrast pairs)

Both procedures are Mean Difference wearing different clothes, and lining them up shows what changed between 2023 and 2025. CAA's construction reads a single, fixed token position — the appended answer letter of a multiple-choice question — from a hand-built dataset of hundreds of (prompt, positive completion, negative completion) triples, one dataset curated per behavior (contrastive-activation-addition, §"3 Method", p. 3). Persona-vector computation reads an averaged position across many tokens — every token of a full free-form response — from rollouts the pipeline generates itself under a contrastive system prompt, then narrows with an LLM-judge filter (trait score above 50 / below 50) before averaging, rather than trusting every generated pair by construction (persona-vectors, §"2.2 Extracting persona vectors", p. 3). The shift is from a format engineered to make the contrast trivially clean — differing by one token, so nothing else needs filtering — to a format that has to manufacture cleanliness after the fact, because free-form rollouts vary in far more ways than an answer letter does. Persona-vectors never states this as a direct successor relationship; it cites CAA only as one instance of prior work needing bespoke data curation. The correspondence between what CAA fixes structurally (one token, pre-guaranteed contrast) and what persona-vector computation fixes statistically (many tokens, post-hoc filtered contrast) only resolves into a legible before/after once persona vectors exists to complete the pair.