I have language model (LM) that produces a probability of responses A and B given a context C: P(A|C),