My experience with this kind of numeric data munging is that small errors like this can totally destroy the final outcome, even without considering the feedback

My experience with this kind of numeric data munging is that small errors like this can totally destroy the final outcome, even without considering the feedbacks between small errors adding up. The article is diligent in showing the impact of the bugs. The individual impacts seem small, but the fact they found so many instances in one little function is beyond worrying. For a little quantification, let’s extrapolate the number of such benign bugs (I make it about 1 per 30 LOC), to get about 500 in the whole codebase. Suppose each bug, charitably, has a 1/1000 chance of catastrophically distorting the result. The chance of no such bugs is a mere 60%. For an imho more realistic probability of 1/100, the chance drops to 0.7%. If the code gives the right answer, it is by pure coincidence.