Just because your code is mechanistic and non-fuzzy doesn't mean that the problem it solves is not fuzzy.

It's about fuzziness and uncertainty of inputs.

There are tolerances of software.

If you have 100 inputs that are exactly the same, and then 1 more thing that is wildly different, how much should you deoptimize?

How flexible your pipeline should be is similar to a multi-armed bandit problem.