Preventative prompting — approximates → Preventative steering

Preventative prompting prepends a trait-eliciting system prompt to every training sample instead of injecting the persona vector into activations, but runs on the same underlying logic as preventative steering: by proactively supplying the undesired behavior during training, it 'cancels out' the pressure that would otherwise push the weights toward that behavior (persona-vectors, §"J.7.2 Preventative prompting on a fact-acquisition task", p. 49). The fact-acquisition case study places the two side by side directly: preventative prompting matches preventative steering at a moderate coefficient (0.5) on hallucination score, new-fact accuracy, and MMLU accuracy alike, but steering at a higher coefficient (1.25) suppresses hallucination further, down to baseline level, while prompting has no comparable way to be made 'stronger.' This exposes a control-range gap between the two operationalizations of the same idea: the activation-space lever is continuously tunable by scaling a coefficient, while the prompt-space lever is close to binary - a system prompt is either present or absent, and its wording can be varied but not smoothly intensified in the way a scalar multiplier can. Preventative prompting therefore demonstrates that the vaccine idea does not strictly require touching activations at all - text conditioning during training reproduces much of the effect - but preventative steering remains the more flexible instrument precisely because 'more steering' is a well-defined operation in a way that 'more prompting' is not.