So what again is the difference for instance against testing scientific codes on a "data-level"? For instance the way how the popular scipy/numpy python package

So what again is the difference for instance against testing scientific codes on a "data-level"? For instance the way how the popular scipy/numpy python packages are tested: https://numpy.org/doc/stable/reference/testing.html#creating-many-similar-tests https://numpy.org/doc/stable/reference/testing.html#creating... (with assertions checking two numerical arrays for similarity) or for instance https://docs.scipy.org/doc/scipy/reference/generated/scipy.fft.fft.html#scipy.fft.fft https://docs.scipy.org/doc/scipy/reference/generated/scipy.f... (Example section where the output of an exemplaric fft run is printed as a 1D array which can be automatically checked by means of https://docs.python.org/3/library/doctest.html https://docs.python.org/3/library/doctest.html). Just to play the devil's advocate: Their ASCII representation of 1d data is error prone and violates minimalistic principles. It is nice to view at and surely easier to read compared to a binary digit sequence, but has little advantage compared to present day testing infrastructure.