I'm sure they check that the matrix doesn't contain a string, or that you're not mixing 32-bit and 64-bit floats, but checking user-defined units as in the arti

I'm sure they check that the matrix doesn't contain a string, or that you're not mixing 32-bit and 64-bit floats, but checking user-defined units as in the article, where each row or column may have a different one seems almost as hard in a dynamically-typed language, unless you do the check at the very bottom level. And that would be unimaginably slow.