Post-hoc (inference-time) steering mitigation — inverts into → Preventative steering

The two mitigations apply the identical update rule to the identical vector with both its sign and its timing reversed. Post-hoc steering subtracts a scaled persona vector at every decoding step after finetuning is already done, h_l <- h_l - av_l, suppressing a trait already baked into the weights (persona-vectors, §"5.1 Post-hoc steering mitigates behavioral shifts", p. 7). Preventative steering adds the same vector during finetuning itself, h_l <- h_l + av_l, steering toward the undesired direction so the model never needs to encode it (persona-vectors, §"5.2 Preventative steering limits behavioral shifts during finetuning", p. 8). The sign flip is not arbitrary: subtracting after the fact fights an already-learned behavior, while adding during training pre-empts the objective's pull toward that behavior, letting the intervention 'cancel out' the pressure rather than override its result. This also creates an asymmetry in when each is usable. Post-hoc steering is reactive - it can be applied to any finetuned model once an unwanted shift is observed, with no need to anticipate anything in advance. Preventative steering is proactive - it requires deciding, before finetuning starts, which trait direction is at risk and steering toward it throughout training, which only works if the risk is known ahead of time. The pair is best read as opposite deployment points along the same intervention, not as competing techniques for the same problem.