The problem with a failed experiment, at least in machine learning, is that it's not always clear what caused the failure. Was it not enough training, was there

The problem with a failed experiment, at least in machine learning, is that it's not always clear what caused the failure. Was it not enough training, was there some small "trick" the model could have used, were the hyper parameters off etc. There are an infinite set of configurations that could fail, and it's not sufficiently useful to know they failed without understanding why they failed. And analysing failures in a useful way is an extremely difficult and fundamental problem. On the other hand, a successful experiment is an extremely rare event and hence interesting by itself.