Three levers on one model: prompting, finetuning, steering — has its alternatives fused in → Steering grows out of inference time

The 2023 paper's three-way comparison was built to answer a choice: for a fixed behavior and a fixed model, is it better to prompt, finetune, or steer? Part of steering's case rested on cost -- generating a CAA vector took forward passes only, under five minutes on one GPU, against roughly ten minutes on two GPUs to finetune the same model on a thousand examples (contrastive-activation-addition, §"6 Comparison to finetuning", p. 6). Preventative steering forfeits exactly that advantage by design: because it is applied during finetuning rather than instead of it, it inherits the full training run's cost and adds a steering intervention at every step on top, never claiming to be cheaper than the finetuning it accompanies. What replaces the cost argument is a change in the question being asked. 2023 asked which lever wins when any of the three could be chosen freely; 2025's preventative pair asks what to do when finetuning cannot be avoided -- because the model is being updated for some other reason, domain adaptation, capability improvement, a safety patch -- and an undesired trait shift is a side effect rather than the whole point. Preventative prompting inherits the same shift in framing: a system prompt eliciting the trait, prepended to every training example rather than offered as an alternative to training, only makes sense once finetuning itself is the fixed constraint being worked around, not one of three competing options.