IMO, go back to 1.0. Adjust the distribution on running plays so they're not so deterministic (i.e. increase the variability). When that's done, start adding bells and whistles to your heart's content as long as they're vetted on the beta and added one at a time.
Even though norbert insists this compartmentalized, step by step system is better, the results are proving otherwise. It's also much more difficult to make adjustments to individual variables since you can't independently test each one on it's own the way the engine is set up. This is also being proved out by the results since the adjustments take so long and seem to have unintended side effects which shows a lot of unknown linkages.