Modelling is perceived as being one of the only tools available to address the new agenda of ecosystem management. However, little is currently understood with regard to the influence of model structure and configuration on predictions and hence management recommendations.In the present study we used a detailed Ecopath with Ecosim (EwE) model of the Barents Sea to test the impacts of food-web aggregation and the removal of weak linkages. Aggregation of a 41-compartment food-web to 27 and 16 compartment systems, greatly affected system properties (e.g. connectance, system omnivory, ascendancy), and also influenced dynamic stability. Highly aggregated models recovered more quickly following disturbances (a pulse of increased fishing pressure) compared to the original disaggregated model.Models aggregated with emphasis placed on particular parts of the food-web (fish, marine-mammals or invertebrates) exhibited marked differences in system indices, despite having the same number of compartments. Models in which invertebrates and basal materials (primary producers and detritus) were heavily aggregated proved particularly resilient to system disturbances. Models focusing on marine-mammals (but in which all other groups were heavily aggregated) also proved very resilient to disturbance, partly due to the slow turnover rates and low biomasses of these top-predatory consumers compared to all other functional groups in the model. Thus, the psychology and decisions of scientists constructing the model can greatly affect its performance and predictions.The Pareto c index is proposed, as a useful measure of skewness towards weak trophic links in food-web models. The 41-compartment ‘control’ model exhibited the highest Pareto c value, and hence was most skewed. Removal of weak links from the food-web, either by eliminating fluxes below a certain threshold or by random-sampling the diet-composition matrix, resulted in models with much lower connectance and Pareto c values. Such models were inherently less stable than the 41-compartment ‘control’ model. Recovery to within 10% of starting values took longer when links had been removed, and the magnitude of fluctuations following a disturbance was also increased.Our findings infer a clear contradiction. Aggregated models possessed fewer weak links but recovered from a disturbance more quickly than disaggregated models (i.e., they were more stable). By contrast, food-webs from which weak links were specifically removed were the least stable of all the models tested. Thus whether weak links are removed through ‘lumping’ or ‘chopping’ seems to have very different system consequences.