Preventative steering — cancels out → Finetuning shift (activation-shift metric)

Finetuning shift is the projection of (post-finetuning minus pre-finetuning) average last-prompt-token activation onto a persona direction, and it strongly predicts post-finetuning trait expression (r = 0.76-0.97) (persona-vectors, §"4.2 Activation shift along persona vector predicts trait expression", p. 6). Preventative steering targets exactly this mechanism rather than the trait score directly: by injecting a*v_l into the residual stream at every training step, it externally supplies the activation movement the objective would otherwise have induced through weight updates, so the model has less need to independently push its own activations along that direction to fit the training data (persona-vectors, §"5.2 Preventative steering limits behavioral shifts during finetuning", p. 8). The measured finetuning shift after training consequently stays small even when training data would normally have driven it up. This gives the metric a second life beyond diagnosis: having been introduced to explain WHY trait expression changes after finetuning, it becomes the paper's own check that a mitigation worked, most directly in the benign-data control, where preventative steering applied to a dataset that would not have shifted the persona anyway leaves both trait score and MMLU accuracy essentially unchanged (persona-vectors, §"J.6 Assessing side effects of preventative steering on benign data", p. 48). The intervention and its evaluation metric share one underlying quantity - activation movement along the persona direction - which is why the fit between them is closer than between preventative steering and the trait-expression score alone.