Ask HN: “Cheating” with boolean rules in machine learning pipeline

A while back there was an informative thread about production ML pipelines - https://news.ycombinator.com/item?id=13821217 I didn’t see any discussion about using hand-crafted rules to tweak the outcome where the ML algorithm wasn’t discriminating well based on patterns in the data. In some cases a boolean rule might fix the confusion, especially if the rule is an actual business requirement, not an ad hoc tweak. For example, I made a classifier to categorize mixtures of chemicals. For each type of mixture, we had a hazardous class and non-hazardous class (based on regulatory status). The hazardous and non-hazardous mixtures could be basically the same stuff, so the classifier often was confused between the two classes. A boolean rule fixed the confusion perfectly. So, practitioners, do you “cheat” in your pipelines? How do you do it?