Preventative prompting — repurposes → System-prompting

Elsewhere in the paper, system prompting is strictly an inference-time lever: it elicits trait expression to validate projection-based monitoring (persona-vectors, §"3.3 Monitoring prompt-induced persona shifts via projection", p. 5), and it is applied to an already-finetuned model as a mitigation baseline, prepending a negative instruction to push behavior back toward normal (§"J.2 Comparison with system prompt-based mitigation", p. 45). Preventative prompting takes the identical mechanism - conditioning a response via prepended instruction text - and relocates it inside the training loop: the same style of trait-eliciting system prompt is prepended to every training example before the loss is computed, rather than to a query at inference (§"J.7.2 Preventative prompting on a fact-acquisition task", p. 49). This changes what the prompt is doing causally, not just when it happens. At inference, a system prompt competes against a fixed, already-trained set of weights, trying to overpower behavior that has already been encoded. As a training-time device, the prompt is present while the weights are still being formed, so the same textual lever changes what the optimization needs to internalize in the first place rather than contesting something already learned. System prompting was built and validated in this paper purely as a way of talking to a finished model; preventative prompting shows the identical tool can instead be pointed at the training process itself, with a materially different causal role despite reusing the exact same surface mechanism.