This is a very under-discussed problem, especially in the context of Data Science. Kudos for a great article. It would have been even better if there were some specific examples in each of the sections describing what can go wrong and how to prevent it. E.g. I worked on a project about bank document processing. The client and the developers had very different ideas about accuracy. The clients wanted the program to run perfectly on 80% of the documents and the developers were thinking that the program should give 80% performance on each document.