I interpreted it more that, in domains where correctness is vital (rather than "good enough"), you want more than one implementation no matter what languages you use.

Maybe that's not what parent was going for, but I think it's like the reproducible/replicable difference in research... can you use the author's code and data, getting the same result... can you use the author's algorithm/pseudocode and data, and get the same result... can you use the author's algorithm/code and different data, and get an _equivalent_ result?