Human societies urgently need more accurate predictions of how the biosphere is going to change under plausible future scenarios. The current major source of uncertainty in these predictions is how the biotic components will interact with a changing climate. Unfortunately, models of the terrestrial carbon-climate feedback make widely diverging predictions even under the same or similar climate change scenarios. The reasons for these different predictions have never been fully understood because of the technical and computational obstacles preventing the intercomparison of model components and structures. We therefore developed a new global vegetation model with the objective of establishing it as a benchmark on which to justify further refinements. We coupled every component process in our vegetation model with global datasets and used Bayesian inference to find the probability distributions for all model parameters. Using various experiments we identified model components and datasets responsible for the major sources of uncertainty. We then simulated our model whilst accounting for full parameter and structural uncertainty to quantify the importance of such uncertainty for climate change predictions. Finally we have wrapped up our study within a computational methodology designed to facilitate objective model intercomparison and rapid model refinement for use by the wider scientific community.