We explore how non-monotonic reasoning can improve a recently introduced approach to parallelization based on non-conservative invasive code refactoring. As the core contribution we suggest to encode parallelization in terms of (libraries of) parallelization strategies -generic defeasible sequences of typed artifacts a.k.a. parallelization recipes. We also suggest to subject these strategies to a specific form of non-monotonic reasoning. This is contrasted with previous approaches based on monotonic inferences. While contemporary parallelization seeks to decide whether a fragment of code is parallelizable our refactoring strategies aim at yielding better parallelization optimized across multiple rewrites. While contemporary parallelization reasons about code we abstract to a higher level where facts about parallelization are accumulated and used as reasoning artifacts. As a proof of concept we illustrate how the suggested generic approach helps optimize data distribution. While we build our discussion around the problem of parallelizing legacy software we also show how generic weaving (c.f. universal) can be introduced with no loss in reusability.