@kareemcarr Generally speaking, this assessment seems about right. I view ML as an applied science where shortcuts get taken due to limitations of time or resourcing (or, dare I say, expertise). In most real-world applications, it's either too difficult or too time-consuming to achieve super high confidence, so we go with the best signals and models available to us.

On my team and with partners, I try to foster a healthy push-pull between the folks who are content with "good enough" and the folks who push for greater rigor before a model goes live. While I don't want to waste time and resources, I also don't want us putting out crappy models, so IMO it's good to have the debate.