Probabilistic models are useful in the analysis of system be-
haviour and non-functional properties. Reliable estimates
and measurements of probabilities are needed to annotate
behaviour models in order to generate accurate predictions.
However, this may not be su cient, and may still lead to
inaccurate results when the system model does not properly
re
ect the probabilistic choices made by the environment.
Thus, not only should the probabilities be accurate in prop-
erly re
ecting reality, but also the model that is being used.
In this paper we propose state re nement as a technique
to mitigate this problem, showing that it is guaranteed to
preserve or increase the accuracy of the initial model. We
present a framework for iteratively improving the accuracy
of a probabilistically annotated behaviour model with re-
spect to a set of benchmark properties through iterative
state re nements.