Despite the difficulties of comparing model predicted variables and observational data (see model Performance), we generally expect that as more data are made available it should be possible to refine our model representations and reduce the uncertainties in model parameters and predictions. Model calibration is, itself, a form of conditioning of model parameters and uncertainties. This will be self-evidently true if a model is a true representation of the system, and the major sources of uncertainties are those associated with the input and boundary condition data, together with observation measurement error. It is less clear that this will be case if the representation of the system is subject to model structure and incommensurability effects. What might still hold in the latter case is that more data will tend to either confirm that the model representation (or, possibly, different model representations) is still acceptable, or alternatively can be rejected.

In either case, the availability of additional data will allow further conditioning of the modelling process that may, hopefully (but not necessarily), result in a reduction in the uncertainty in predictions. Since such prediction uncertainties enter directly into the estimation of risk, there is also the possibility of conditioning risk estimates. In the case where formal statistical likelihood measures can be assumed to be acceptable, this conditioning process can be carried out using either analytically or within a Bayesian Likelihood framework. In the case where informal or fuzzy model evaluation measures are used, it will still be possible to combine prior likelihood measures with those arising from an evaluation against the new data in some way (using Bayes equation, Fuzzy Union, or other combination method) to produce a posterior likelihood or ranking of different models. In this informal framework, however, the result cannot be considered as a statistical probability of estimating an observation conditional on the model.