The preconditions of an API method are constraints on the states of its
receiver object and arguments intended by the library designer(s) to
correctly invoke it in the client code. There have been two main kinds of
approaches for automatically inferring API preconditions. The first kind of
approaches mines the frequently checked conditions guarding the API usages in
the client code and generalize them into preconditions. The second kind of
approaches analyzes the implementation of the API to compute preconditions.

In this paper, we report an observation that the usage-based approach often
produces preconditions stronger than those intended while the implementation-based
produces weaker ones. Our finding calls for a new direction of
integrating those kinds of precondition inference approaches and refinement
solutions to reduce the differences between sets of inferred preconditions.