A computational study is performed in which the predictive capabilities of a range of eddy-viscosity and second-moment-closure models are examined by reference to a separated flow behind a backward-facing step in an expanding channel. The models include three second-moment-closure variants, all being of the 'Launder-Reece-Rodi' type, two RNG k-epsilon forms, one combining the RNG approach with a non-linear eddy-viscosity formulation, and a low-Re k-epsilon model. The study demonstrates that to achieve a solution similar to that returned by second-moment closure, the RNG formulation needs to be implanted into a non-linear eddy-viscosity framework; neither returns, on its own, the correct behaviour, not even for mean-flow features. Moreover, relatively minor variations within second-moment closure-specifically, such relating to wall-induced effects on turbulence isotropisation and to stress diffusion-can significantly alter the overall performance of the closure. All models specifically designed to return realistic solutions for normal stresses seriously over-estimate anisotropy.