After demonstrating that the preliminary model predicted testpoints significantly better than random, we modelled the species’potential geographical distribution using all available localities.Because GARP is an artificial-intelligence application withstrong stochastic elements, it produces no unique solution. Totemper among-model variation, we made three models perspecies and developed a composite prediction for each species:any pixel where the species was predicted present by at leasttwo of the three models was considered predicted present. Allfurther analyses were based on these composite models