Quite often, in weakly consistent implementations of replicated data types (CRDTs), the notion of correctness is subtle. For example, in a weakly consistent bank account, it might be acceptable for a getBalance operation to not return accurate balance, but it is not acceptable for it to return a negative balance. Likewise, it may be acceptable for a weakly consistent shopping cart to miss showing items added in a different user session, but it is not acceptable for it to miss items added in the same session. Recent literature [Sivaramakrishnan et al, PLDI'15] proposes a  specification language for consistency properties that is sensitive to such subtleties in correctness specification. However, crafting a specification that concretizes the notion of correctness is quite challenging, unless concrete manifestations of incorrectness (anomalies) are known. On the other hand, programmers regularly write test cases that more-or-less capture the desired semantics of the application. In this talk, I describe how such test cases can be used to automatically infer consistency requirements of an application, thus obviating the need for programmer annotations.