Context and context-awareness have been central issues in ubiquitous computing research for the last decade. Advances with regard to context acquisition and activity recognition allow interesting small scale applications. However in larger systems including many sensors and actuators and spanning multiple administrative domains are still remain as unsolved central issues. Particularly, in the areas of reasoning and context-fusion there are many open questions. In this paper we motivate large scale context-aware systems that have support for the full life cycle using a prototype that was implemented and tested. We then describe in detail the underlying architecture which supports dynamic context-aware systems and includes mechanisms for context fusion and reasoning. Also, we propose a new approach for evaluating context-aware systems. The approach is an adapted expert evaluation; well known in the user interface domain, but using a carefully selected set of heuristics specifically targeted at context-aware systems.