> you could do substantially the same thing by just prompting/LoRA finetuning the model to produce a single-token output ("yes" or "no")

You could probably achieve this with logit masking. Or equivalently, comparing the "yes" vs "no" logprobs in the final dis-embedded vector.