A novel uncertainty assessment methodology, based on a statistical non-parametric approach, is presented in this paper. It achieves quantification of code physical model uncertainty by making use of model performance information obtained from studies of appropriate separate-effect tests. Uncertainties are quantified in the form of estimated probability density functions (pdfs), calculated with a newly developed non-parametric estimator. The new estimator objectively predicts the probability distribution of the models error (its uncertainty) from databases reflecting the models accuracy on the basis of available experiments. The methodology is completed by applying a novel multi-dimensional clustering technique based on the comparison of model error samples with the Kruskall-Wallis test. This takes into account the fact that a models uncertainty depends on system conditions, since a best estimate code can give predictions for which the accuracy is affected by the regions of the physical space in which the experiments occur. The final result is an objective, rigorous and accurate manner of assigning uncertainty to coded models, i.e. the input information needed by code uncertainty propagation methodologies used for assessing the accuracy of best estimate codes in nuclear systems analysis. The new methodology has been applied to the quantification of the uncertainty in the RETRAN-3D void model and then used in the analysis of an independent separate-effect experiment. This has clearly demonstrated the basic feasibility of the approach, as well as its advantages in yielding narrower uncertainty bands in quantifying the codes accuracy for void fraction predictions.[All rights reserved Elsevier].