Response-conditioned contrastive vector computation — filters its rollouts by → Trait expression score (LLM judge)

Not every rollout generated under a contrastive system prompt actually displays the trait it was prompted to display, and computation discards the ones that don't before any activation is averaged. For each of the 400 rollouts generated per trait (20 extraction questions times 10 rollouts times 2 conditions), the pipeline scores the response with the freshly generated rubric and keeps only those exceeding 50 for positive-prompted rollouts or falling below 50 for negative-prompted ones, discarding the rest (persona-vectors, §"2.2 Extracting persona vectors", p. 3). This matters because the contrastive system prompt only sets an intention, not a guarantee: a model instructed to act 'evil' can still produce a mild or off-target response, and averaging such a response in with genuinely trait-exhibiting ones would blur the mean-difference signal with noise from responses that never actually diverged along the target direction. The threshold is symmetric and fixed at the scale's midpoint rather than tuned per trait, treating 50 as the boundary between 'exhibits' and 'does not exhibit' regardless of which of the three main traits or four appendix traits is being extracted. The filtering step is where trait-expression-score's role shifts from a downstream outcome measure, which is how the rest of the paper mostly uses it, to an upstream data-quality gate inside the extraction pipeline itself.