Hello there.
What should we optimise today?

Each prompt is rewritten by a reinforcement-learning policy before it reaches the model. You see both versions, then both answers are scored so you can compare efficiency before and after.