Sure, there are of course runtime checks you can do (and I now do after having been bitten by this before), but that’s not necessarily better than having a static analyzer which can guarantee correctness without executing a single line of code.
Yes, the example @beisner gave here (the val loop crashing after the training is done) is kind of a bad example, and some frameworks (like Pytorch lightning) do do a "full workflow check" before going into training, but his overall point that variadic generics could massively impact dev/researcher productivity stands, I think.