Software architecture degradation is a phenomenon that frequently occurs during software evolution. Source code anomalies are one of the several aspects that potentially contribute to software architecture degradation. Many techniques for automating the detection of such anomalies are based on source code metrics. It is, however, unclear how accurate these techniques are in identifying the architecturally relevant anomalies in a system. The objective of this paper is to shed light on the extent to which source code metrics on their own can be used to characterize classes contributing to software architecture degradation. We performed a multi-case study on three open-source systems for each of which we gathered the intended architecture and data for 49 different source code metrics taken from seven different code quality tools. This data was analyzed to explore the links between architectural inconsistencies, as detected by applying reflexion modeling, and metric values indicating potential design problems at the implementation level. The results show that there does not seem to be a direct correlation between metrics and architectural inconsistencies. For many metrics, however, classes more problematic as indicated by their metric value seem significantly more likely to contribute to inconsistencies than less problematic classes. In particular, the fan-in, a classes’ public API, and method counts seem to be suitable indicators. The fan-in metric seems to be a particularly interesting indicator, as class size does not seem to have a confounding effect on this metric. This finding may be useful for focusing code restructuring efforts on architecturally relevant metrics in case the intended architecture is not explicitly specified and to further improve architecture recovery and consistency checking tool support.