lol, predict it will always predict bro
but the issue is accuracy metrics that you are not taking in consideration. For example, once the main neural model is trained, we can use its frozen representations at certain layer and learn a linear classifier on it by probing it. This essentially checks if the representation in the given layer has enough information for the specific task.
What I mean here it's that round starts provide not enough information, and the model will likely predict a 50-50 win chance. Whereas rounds that are deeply disbalanced in terms of fire power, you will see AI predicting the obvious likely winner (e.g. team with more guns/nades etc). Thus, it's way more interesting the modelling of post-plant situations in which small details can be way more impactful. In other words, post-plant actions will produce larger perturbations which will have, on average, a larger effect on the neural network prediction.