(Left) In stage 1, problematic regions are predicted using a newly developed error predictor that looks for local strain in the model and poor local density-fit. These selected regions are subject to iterative fragment-based rebuilding within a Monte Carlo sampling trajectory. Refinement in this stage is restricted to using one-half of the data, referred to as the training map. (Middle) In stage 2, the best models from the ~5000 independent Monte Carlo trajectories are selected. Models are selected based: on agreement to the validation map (independently constructed from the other half of the data), then by model geometry as assessed by MolProbity, and finally, on agreement to the full reconstruction. At this point, the selected models should in general have good fit-to-density and good geometry without overfitting to the data. (Right) In stage 3, using the 10 best models selected, we then optimize against the full reconstruction. Two half maps are used to choose the optimal density weight to refine structures using full-reconstruction. Finally, these top 10 models are optimized (without large-scale backbone rebuilding) into the full-reconstruction, which alternates with voxel-size refinement iteratively. Finally, these models are subject to B-factor refinement.