Reward Prediction Normalization — fixes the underdetermined scale of → Reward model (RM)

A reward model trained only on comparisons never sees an absolute number, so its output's zero point and scale are free parameters left unconstrained by the training objective (deep-rl-human-prefs, §"2.2.1 Optimizing the Policy", p. 5). Reward normalization pins both down by rescaling the predictor's output to zero mean and a fixed standard deviation before it reaches the RL algorithm. That fixed standard deviation is chosen per domain specifically so existing RL hyperparameters could be reused unchanged: 1 for the MuJoCo tasks and 0.05 for Atari, a choice the paper says let it keep the same entropy bonus and learning rate as its true-reward baselines rather than retune them for an arbitrarily-scaled learned reward (deep-rl-human-prefs, §"A.2 Atari", p. 15). Without this step, the same reward model, valid up to any positive affine transformation of its output, could silently break RL hyperparameters that were tuned assuming a particular reward magnitude.