Open models as subjects, frontier models as instruments — supplies the instruments that run → From trait name to vector: extraction becomes automatic

The subject/instrument split this theme documents at the level of the whole experimental stack is not incidental to the extraction pipeline -- it is required by what a persona vector is. Stage one, artifact generation, is entirely instrument work: Claude 3.7 Sonnet writes the five contrastive system-prompt pairs, forty questions, and rubric from nothing but a trait name and description (persona-vectors, §"2.1 Generating trait-specific artifacts", p. 3). Stage two cannot be instrument work at all: the ten rollouts generated under each system prompt have to come from whichever model -- Qwen2.5-7B-Instruct or Llama-3.1-8B-Instruct -- the vector is being built for, because the vector is a direction in that specific model's own residual stream, not a portable artifact Claude could hand over pre-made. GPT-4.1-mini's trait expression score sits in between: it never touches either model's activations, only the text both produce, filtering stage-two rollouts by a threshold (score greater than 50 for positive prompts, less than 50 for negative) before extraction happens (persona-vectors, §"2.2 Extracting persona vectors", p. 3). So the pipeline does not just use the instrument stack -- it is organized around the one constraint that stack imposes: text can be outsourced to a frontier model at any stage, but the activations a vector is made of cannot, which is why stage two is the only stage that has to run on the subject.