Preventative steering — outperforms → Train-time projection-regularization baseline

The regularization baseline penalizes change in the projection of activations along a trait direction during finetuning - an intuitive alternative to steering that targets the same quantity finetuning shift measures. It reduces the projection shift somewhat but the model still expresses the trait after training, because optimization pressure finds it cheaper to represent the trait through alternative activation-space directions the loss term does not penalize, rather than genuinely not learning it (persona-vectors, §"J.5 Train-time regularization along undesirable persona directions", p. 47-48). This is a Goodhart-shaped failure: the regularizer targets one measured proxy for trait presence, and gradient descent routes around exactly that proxy while still solving the underlying prediction problem the trait-eliciting data poses. Preventative steering avoids this failure by a different mechanism than merely discouraging movement along the direction - it actively supplies that movement externally at every step (§5.2, p. 8), which removes the training pressure to encode the trait via any direction, not only the one being monitored. Where regularization fights the symptom (the projection value it can see), steering removes the underlying driver (the need to fit trait-eliciting responses by changing internal representations at all). The comparison shows that penalizing a measured direction and preempting the need for a direction are not equivalent interventions, even though both operate on the same projection.