Since the model doesn't get to observe the actual situation being reported on, such an improvement in accuracy would only be random chance and should not be rewarded.