It seems the purist definition of solving protein folding means specifying the sequence of actions to arrive at the native structure for a given input sequence. This implies your solution is grounded in a physical reality. Such approaches (i.e MD simulations) have not been very successful to my knowledge and are computationally intractable for most proteins. Even if you are able to do this, how do you verify that your folding mechanism is in fact correct? We don't have ground truth experimental data at the temporal resolution one would want to verify such a prediction.
For now, the only practical framing of the protein folding problem is to predict a native structure given an input sequence. This is simply because we have the requisite data...