The benchmark exercise discussed in this paper was conducted within the OECD/NEA project HYMERES. The specific experiment in the PANDA facility chosen for the present benchmark addresses the stratification erosion in a vessel where the upper region contained initially a mixture of steam and helium, and the remaining volume was filled with steam. The mixing is induced by a vertical steam jet, which originates from the exit of a circular pipe located below the bottom of the helium-rich layer. The stratification erosion process is somewhat slowed down by a small circular plate above the jet source. The exercise consisted of a blind phase, and an open phase. Two sets of blind simulations were requested: one set obtained using a "common model", and a second set produced by a "best estimate" model. For the "common model", a list of recommendations was given, whereas for the "best estimate" model, each participant was free to choose the modelling approach. The submitted results for the erosion times were in a large band, and especially the large differences in the results with the "common model" were not expected. The results of the best estimate simulations showed that the combination of mesh and modelling approach can lead to a wide spread of results. The most important difficulty in interpreting the results and finding the reason of the large deviations was the lack of information on the velocity field downstream of the obstruction. Therefore, for the open phase extended data from auxiliary, "zero" tests (for similar conditions but without helium layer) were provided to the participants to permit a more basic validation of their models, using a "multi-step approach". The step-by-step validation permitted some progress with respect to some of the items identified in the blind benchmark. However, large discrepancies with data in the final analyses of the test are observed, which cannot be easily attributed to specific model deficiencies or insufficient detail of the mesh. These results raised some questions in relation to best practice guidelines for the use of Computational Fluid Dynamic (CFD) codes for containment analysis and indicated needs for further CFD-grade experiments.