Here is the simple version:
Until now, we had been scoring based on RMSD. RMSD = root mean square deviation from the correct native structure. What this means is that we KNOW the native structure of the protein, which is has RMSD of 0.0 (it doesn't deviate from itself). Any generated structure is evaluated based on how closely it approaches the native structure, where the score is the RMSD. This method is "cheating" because normally we would be trying to fold proteins whose native structure is unknown - the folding will strive to find the structure which os closest to the native protein.
Thus, we will move to scoring by using energy-based scoring functions. In this case the protein's score will be based on its energy and compactness, without relying on the native structure for comparison. The idea remains that lowest-scoring proteins should most closely resemble the natural-occuring structure. However, in this case, the prediction is ab-initio, or in layman's terms, from scratch.