You are correct that they rely on a random seed, a fixed model, and other assumptions that I made. If you want to generalize it without these assumptions, you can just look at the output differently, i.e., it can be the length of the output, type of the output, etc.
In order to prevent degradation, you have to monitor your model and everything surrounding it. check my article on monitoring: https://towardsdatascience.com/monitor-stop-being-a-blind-data-scientist-ac915286075f
Lastly, in order to test data, use packages such as TDDA.