Third party software is now increasingly used in systems with high dependability requirements. This evolution of system development raises new challenges, in particular regarding the implementation of fault-tolerance. As systems are often built of black-box components, some crucial aspects of their behavior regarding replication cannot be handled. This is also true to some extent for open-source components as mastering their internal behavior is sometimes very tricky (e.g. OS and ORBs). During the last decade reflection has emerged as a very fruitful paradigm for the disciplined management of non-functional aspects, among which fault-tolerance. In this paper we discuss how to apply reflection to multi-layer systems for implementing fault-tolerance in an independent and principled manner. We analyze the connections between the underlying assumptions of fault-tolerance strategies and the different layers of a system. Based on this multi-layer analysis we show how the requirements of a family of replication algorithms can be addressed on a concrete architecture, resulting in what we name Multi-Layer Reflection.