What's wrong with the plan, and what would you do instead?

1. ANothing: a Bradley-Terry score is a calibrated measure of answer quality.
2. BThe bar should be 0.5, since the model outputs a sigmoid probability.
3. CScores are fixed only up to a per-prompt shift; compare within a prompt.
4. DRetrain the reward model with MSE on the pairs so all scores share a scale.