Cyber-physical systems (CPS) are systems of cooperating autonomous components which closely interact with and control the physical environment. Being distributed and typically based on periodic activities, CPS have to cope with the problem that data capturing a distributed state of the system and its environment are inherently inaccurate (they represent belief on the state). In particular, this poses a problem when dependability is being pursued. In this paper we address this issue by modeling belief at the architecture level. In particular, we enhance the architecture by models describing belief inaccuracy over time. We exploit these models to quantify at runtime the impact of belief staleness on its inaccuracy. We then use this quantification to drive architectural adaptation with the aim to increase dependability of the running CPS system.