A model of software complexity and reliability isdeveloped. It uses an evolutionary process to transitionfrom one software system to the next, while complexitymetrics are used to predict the reliability for eachsystem. Our approach is experimental, using datapertinent to the NASA satellite systems applicationenvironment. We do not use sophisticatedmathematical models that may have little relevance forthe application environment. Rather, we tailor ourapproach to the software characteristics of thesoftware to yield important defect-related predictors ofquality. Systems are tested until the software passesdefect presence criteria and is released. Testingcriteria are based on defect count, defect density, andtesting efficiency predictions exceeding specifiedthresholds. In addition, another type of testingefficiency—a directed graph representing thecomplexity of the software and defects embedded in thecode—is used to evaluate the efficiency of defectdetection in NASA satellite system software.Complexity metrics were found to be good predictorsof defects and testing efficiency in this evolutionaryprocess.