Researchers and practitioners sometimes presume that using a previously "validated" instrument will produce "valid" scores; however, contemporary views of validity suggest that there are many reasons this assumption can be faulty. In order to demonstrate just some of the problems with this view, and to support comparisons of different observation protocols across contexts, we introduce and define the conceptual tool of an "observation system". We then describe psychometric evidence of a popular teacher observation instrument, Charlotte Danielson's Framework for Teaching, in three use contexts--a lower-stakes research context, a lower-stakes practice-based context, and a higher-stakes practice-based context. Despite sharing a common instrument, we find the three observation systems and their associated use contexts combine to produce different average teacher scores, variation in score distributions, and different levels of precision in scores. However, all three systems produce higher average scores in the classroom environment domain than the instructional domain and all three sets of scores support a one-factor model, whereas the Framework posits four factors. We discuss how the dependencies between aspects of observation systems and practical constraints leave researchers with significant validation challenges and opportunities.