Besides being The Problem of our times, detecting misinformation is also a fascinating engineering challenge. It's non-stationary, has no clear definition, and labels are mostly context-dependent. Misinformation *datasets* are none of these things. This article is dedicated to that dichotomy.