Dear phenix.refine community:
 
I am solving a protein structure using experimental phasing. In order to do that I combined two derivative isomorphous datasets into a double SAD experiment (to increase the individual weak anomalous power from each dataset) using SHARP.  My resolution is 3.2 A, and the space group is C2221.
 
I managed to build a model using the density-modified phases from SHARP. Then I started refining against one of my derivative datasets because my native dataset is non-isomorphous due to the heavy meal soaking. Then I made two mistakes :(
 
1. I did not realize at the beginning of my refinement-building rounds that I was refining against the dataset that showed the worst statistics.
 
2. I was using a dataset with only FP and SIGFP but not DANO and SIGDANO.  I used the phases from my SHARP map (also the experimental phase restraints)
 
Now I did a new run of refinement using my best dataset (better statistics) and also including the anomalous data FP, SIGFP, DANO and SIGDANO. Obviously, my R-work and R-free got pretty much better.
 
Nevertheless, the R-free was better than R-work at the beginning of the run, at the end of the refinement R-free and R-work inverted as usually (being R-free worst than R-work)  r_work = 0.2582  r_free = 0.2888 . Certainly, there was a mistake with a new assignment of the free set of reflections. Now I am aware that the set of free reflections has to be the same all over the refining procedure. 
 
Here are the questions

How can I be sure that my R-free is still a safe cross validation of the correctness of my structure?
 
Is it useful to perform a round of simulated annealing in order to erase any wrong information in my model and start the refinement rounds again?

Do you have any suggestions?
 
All the best
 
Jose Farias