Evaluating steering: behavior scores and capability floors — certifies the dial metaphor for → Seven alignment worries become seven dials

A dial is only a dial if turning it does what it claims and nothing else, and this theme supplies both halves of that proof for the other. The multiple-choice sweep across every layer at multipliers of −1 and +1 shows each of the seven behaviors moving monotonically with the sign of the multiplier on 50 held-out questions, the evidence that a behavior is actually being dialed up or down rather than merely correlated with the vector (contrastive-activation-addition, §"4.1 Multiple-choice question datasets", p. 4). The MMLU capability floor supplies the other half: sampling ten questions from each of 57 subjects and reporting average probability on the correct answer, the paper finds "our intervention does not significantly affect MMLU performance" under the same interventions that move the seven behaviors (contrastive-activation-addition, §"7 Effect of CAA on general capabilities", p. 6). Without both checks, "seven dials" would just be seven unverified vector labels; together they certify that the dial metaphor, and not just the vector's existence, is earned.