Thoughts on this: Do we want to have a bit more logical separation that is conducive to NIVC?

For example, perhaps we can parse the JSON into the machine. Pass the machine as step_out into interpreter stage that grabs the location (indices) of where the intended object is and then pass on this new subarray of the machine into the next iteration of this process.

For example, consider:

{
   "a": {
       "inner_a": { 
           "inner_inner_a": "my_string_value"
        }
    }
}

where we want to extract a.inner_a.inner_inner_a. We parse the whole JSON, pass machine into ExtractValue which outputs data of the start and end of a (and here we can even trim bottom off stack and truncate the machine state further possibly, though let's not get into pre-mature optimization), then this is passed again into ExtractValue, and it returns start and end indices of inner_a as step_out. Finally, we know we are at the final retrieval so we call a new circuit instance in NIVC ExtractStringValue which then hunts within the range output previously for inner_inner_a and gets the String value there.

I imagine the role of array is similar and we can just choose to use ExtractArray in NIVC