System One hasn't said how RLCD works, but they do say it is explicitly training models to output "calibrated" probabilities, which makes it distinct from RLVR. This is how they describe it:
> System One models are trained for calibrated decisions: their probabilities are optimized against outcomes to reflect uncertainty.