People's past experience with the justice system would no doubt be part of the model, as well as factors possibly including: Career area, Dependents/spouse, Time in current job, Past (unconvicted) run ins with cops, Known drug addictions, Track record of arresting cop and sentencing judge, ect.

With a good model, it would be hard to charge "normal" people in a way that actually gamed the statistics, because their probability to re-offend is very low to begin with. When they don't re-offend it would be expected behavior and not represent in drop in observed recidivism vs expected recidivism. So no bonus.

I would expect the lowest hanging fruit to be in drug addicts and thieves, there is a very large body of knowledge about rehabilitating those two groups.These would be the two groups where I expect to see the largest difference between expected recidivism in the current system vs. a treatment group with psych professionals and job training provided.