We find here a distinction between two different kinds of rules: using a Wittgensteinian terminology we might speak of rules governing the working of language games, and the rules governing relations among language games. How shall we define rules defining contents inside different contexts? In order to blend concepts we need concepts to be blended; in order to lift an inference from a context to another we need an inference to be performed. While we are beginning to study a new level of rules among contexts, we run the risk to find inside each context the same problem of semantic holism we have found at the beginning. Do the discovery of new kinds of rules which run across contexts have some impact on this problem? I will come back on these question after some remarks on one last solution on the market: local holism.