At the recent WICSA, a debate was held on the topic "which matters more: the architecture as designed or the one the running system happens to have?"
A bit flippantly, I realize, I remarked in my keynote that this was a silly topic. It's akin to asking what matters more to water, oxygen or hydrogen?
In short, my take is that both matter, depending on what you are trying to do. Remember my personal bias: the primary artifact of any meaningful software development organization is running software - not documentation, not models, not even source code. However, my personal bias continues to suggest that this does not mean these other artifacts are irrelevant - source code is the ultimate human-readable statement of the running system, models represent abstractions of that system from various points of view (most models should be tossed and only some retrained) and documentation serves to capture design decisions not reflected in the code.
That last phrase demands some explanation, for it's the key to why I also suggest that the architecture as designed is important. In every interesting system there exists a software design entropy, meaning that there is a loss of design information from vision to execution. In other words, there are some design decisions that are not directly observable from the running system or its source code. We know how to recover some design patterns automatically from source code; we don't yet know how to recover higher level patterns automatically from source code, largely because a) we haven't codified all the patterns we are trying to recover and b) its a wickedly hard problem to identify the signature of such higher-level patterns from the noise of the source code. For this reason, the architecture as designed is significant, if and only if what we mean by architecture encompasses the significant design decisions that are not recoverable directly from the source code.
This balance is also key to avoiding the over-engineering of a system's architecture. If your design can be expressed in code, do it; if it can't then model it first.