A predictive analysis takes an execution trace as input and discovers concurrency bugs without accessing the program source code. A sound predictive analysis reports no false positives, which sounds like property that can be defined easily, but which has been defined in many different ways in previous work. In this paper, we unify, simplify, and generalize those soundness definitions. Our soundness definition is graph based, separates thread-local properties and whole-execution properties, and works well with weak memory executions. We also present a three-step proof recipe, and we use it to prove six existing analyses sound. This includes the first proof of soundness for a predictive analysis that works with weak memory.