Pervasive systems are characterised by networked heterogeneous devices. To fulfill the security requirements of an application, these devices have to abide by certain policies. However, as the contingent interaction between devices in all possible contexts within evolving pervasive systems devices cannot be known at development time, policies cannot be dedicated to concrete security mechanisms which might later not be supported by the devices present in the network. Therefore, policies need to be expressed at a more abstract level and refined appropriately to suit applicable mechanisms at run time. In this paper we describe how security policies can be combined with ontologies to support such an automated policy refinement. As thereby policy decisions depend on semantic descriptions, the correctness of these descriptions must be verifiable at a later time for policy decisions to be evidential. We therefore propose Trusted Computing-based approaches on generating proofs of correctness of semantic descriptions deployed in policies.