Reference study: Persona Vectors: Monitoring and Controlling Character Traits in Language Models (Chen et al., Anthropic, 2025) — extracts activation directions for traits (e.g. sycophancy, hallucination, “evil” persona), monitors them during generation, and steers behavior by adding/subtracting the vector. Maps directly onto psychscanner’s existing persona trials, using the model internals already exposed by the nnsight/nnterp backends.