While they can be useful tools to support decision-making in ecosystem management, robust simulation models of ecosystems with alternative states are challenging to build and validate. Because of the possibility of alternative states in model dynamics, no trivial criteria can provide reliable and useful metrics to assess the goodness-of-fit of such models. This paper outlines the development of the model TRITON, and presents simulation-based validation and analysis of model sensitivity to input parameters. TRITON captures the local dynamics of seaweed-based rocky reef communities in eastern Tasmania, which now occur in two alternative persistent states: (1) either as dense and productive seaweed beds, (2) or as sea urchin 'barrens' habitat, i.e. bare rock largely denuded of macroalgae and benthic invertebrates due to destructive grazing by sea urchins. Pattern-oriented-modelling, i.e. comparing patterns in model dynamics across Monte-Carlo simulations with direct observations of Tasmanian reef communities over large scales, provides a valuable approach to calibrate the dynamics of TRITON. Using the computationally efficient, model-independent extended Fourier amplitude sensitivity test, we identify fishing down of predatory lobsters, sea urchin recruitment rate, as well as seaweed growth rate as key parameters of influence on overall model behaviour. Through a set of independent sensitivity tests, we isolate different sets of drivers facilitating the 'forward' shift from the seaweed bed to the urchin-dominated state, and the reverse or 'backward' shift from denuded sea urchin barren to recovery of seaweed cover. The model suggests that rebuilding populations of large rock lobsters, which predate the urchins, will be effective in limiting ongoing formation of sea urchins barrens habitat, but that the chances of restoring seaweed beds from extensive barrens are relatively low if management relies solely on rebuilding stocks of large rock lobsters. Moreover, even when