So, phase one deals with outlier identification. Phase 2 would not have to look at the history, as the model in phase one did this. It would basically take the last draw in the 150 draw history chunk and calculate all of the probabilities for each neutral and return the digit with the highest probability.

Of course there would need to be a process in place to prevent ties, perhaps this could be it's last appearance from phase 1.

Still a great deal of design work to do to understand the process, but this is definitely a start. This is all about the last state of the data, row 150, and the probabilities of the entire neutral distribution.