The moment you touch the test data you have tainted it because your brain can do a crude search on the space of solutions and reconcile them after the fact. This is why any preprocessing and cleaning is parameterized on the training set and then applied to the data, that also includes the class of models you can consider and is why you never select final models based on their performance on the test set; it becomes biased.
[1] probability of getting a good model.
[2] how much the empirical performance underestimates the expected performance.
[3] the number of models you can select from a particular class of models.
Unluckily for them their online literacy says I'm too stupid to write simulation code them so their loss.
But imagine the FPGA-grunt in those things!