We compare computational fluid dynamics (CFD) predictions using a steady-state Reynolds Averaged Navier-Stokes (RANS) model with experimental data on airflow and pollutant dispersion under mixed-convection conditions in a 7 x 9 x 11m high experimental facility. The Rayleigh number, based on height, was O(1011) and the atrium was mechanically ventilated. We released tracer gas in the atrium and measured the spatial distribution of concentrations; we then modeled the experiment using four different levels of modeling detail. The four computational models differ in the choice of temperature boundary conditions and the choice of turbulence model. Predictions from a low-Reynolds-number k- model with detailed boundary conditions agreed well with the data using three different model-measurement comparison metrics. Results from the same model with a single temperature prescribed for each wall also agreed well with the data. Predictions of a standard k- model were about the same as those of an isothermal model; neither performed well. Implications of the results for practical applications are discussed.