Multiple-choice contrast-pair format — repurposes the ab mechanic of → Multiple-Choice Comparison Format

Constitutional AI's Multiple-Choice Comparison Format presents a conversation, a constitutional principle, and two labeled candidate responses, then reads the feedback model's answer log-probabilities off as a preference label used to train the preference model (constitutional-ai, §"4.1 Method", p. 10). CAA's multiple-choice contrast-pair format looks superficially identical -- a question ending in an appended answer letter, 'A' or 'B' -- but the letter is not read out as a label at all. Instead, CAA extracts residual-stream activations at the token position of the answer letter and differences them across hundreds of paired prompts to build a steering vector, deliberately designed so the two prompts in a pair differ by only that single token (contrastive-activation-addition, §"3.1 Sourcing datasets", p. 3). An appendix validation shows what the letter is doing internally: conditioned on having 'chosen' A or B, the model's own continuation naturally justifies that choice, meaning the letter functions as a hook into the model's internal state rather than a data point to be collected (contrastive-activation-addition, §"B Answer conditioning leads to behaviorally consistent continuations", p. 13). Same surface mechanic, a year apart, put to two unrelated uses: reading out a preference versus reading out a representation.