Claude Sonnet as pipeline/data-generation tool
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — inherited
Claude 3.7 Sonnet and Claude 4.0 Sonnet are used throughout as generation tools rather than experimental subjects: Claude 3.7 Sonnet generates the pipeline's trait artifacts and the trait-eliciting/EM-like dataset responses, while Claude 4.0 Sonnet generates the interpolated trait-suppressing-to-trait-promoting system prompt sequences used for the monitoring experiments.
A paper about how training data shapes a model's character needs a source for that training data, and here the source isn't a human writer at all but a separate, more capable model doing the generating on request. Two versions of it are used for entirely different jobs across the paper.
One small, single-call task producing a trait's prompts and questions sits alongside one much larger job writing thousands of training responses across every dataset in the paper. The flawed responses in one whole family of datasets rest entirely on this same model's own judgment about what a plausible mistake looks like, with no human review step in between.
Two Claude models, two jobs
Claude 3.7 Sonnet and Claude 4.0 Sonnet appear throughout the paper as generation tools rather than experimental subjects — models whose outputs become the paper's data and prompts, never models whose activations are read or steered. Claude 3.7 Sonnet authors the trait-specific artifacts produced by the extraction pipeline and the response text in the trait-eliciting and EM-like finetuning datasets; Claude 4.0 Sonnet is used separately, to generate the interpolated system-prompt sequences used in the monitoring experiments: "we use Claude 4.0 Sonnet to generate eight prompts that smoothly interpolate between trait-suppressing and trait-promoting instructions" (persona-vectors, §"3.3 Monitoring prompt-induced persona shifts via projection", p. 5).
A small job and a large one
Claude 3.7 Sonnet's role inside Trait-specific artifact generation is a single one-shot call per trait: a generic prompt template feeds it a trait name and description, and it returns 5 contrastive instruction pairs, 40 evaluation questions, and 1 evaluation rubric in one pass (persona-vectors, §"2.1 Generating trait-specific artifacts", p. 3). That is small next to its other job: for the Evil and Hallucination datasets, Claude 3.7 first generates the questions themselves — 50 domains times 100 questions each, roughly 5k questions per trait — and then, for every dataset in the paper, generates three separate response versions (Normal, I, II) per question (persona-vectors, §"D.1 Question collection.", p. 35; §"D.2 Response generation", p. 35). The Sycophancy dataset alone contributes roughly 10k training questions needing three response versions apiece.
Authoring the EM-like datasets' flaws
Every response in the five EM-like datasets — Medical, Code, MATH, GSM8K, and Opinions — is written by Claude 3.7 Sonnet, run at temperature 1.0 with an 8000-token limit and a 2000-token thinking budget, using one of two domain-generic templates: a "Mistake Response Generation" prompt for Medical, MATH, GSM8K, and Opinions, and a "Vulnerable Code Response Generation" prompt for Code (persona-vectors, §"D.2 Response generation", p. 37-38). Each template asks Claude to write a correct response, then a subtly wrong one, then a more severely wrong one, along with its own explanation of what makes each wrong answer wrong; the paper describes no human review step for individual responses, only automated discarding of generations that fail to parse as JSON (persona-vectors, §"D.2 Response generation", p. 38). The entire dataset family used to show that domain-specific errors alone can shift a model's persona therefore rests on one frontier model's own judgment about what a plausible medical, mathematical, or security mistake looks like.