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... (with assertions checking two numerical arrays for similarity) or for instance 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).

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.

I’ve used these sorts of tests before, but under the name of ‘golden’ rather than ‘expect tests’. To complement the post’s OCaml example, here’s an article on how to use them in Haskell: https://ro-che.info/articles/2017-12-04-golden-tests

Yeah, the term "golden" is common in the hardware world, which corresponds to expected signal level traces either captured from a physical device with a logic analyzer, or generated with an ideal model.