RLCD is a calibrated, schema-conditioned extension of pairwise reward modeling: Bradley–Terry becomes Plackett–Luce, and the reward model becomes Jev’s typed decision interface.