I’ve already built arbiter systems by stacking one llm on top of another, the first with specific role of determining a defined variable for number two to use in its process.
Yea I like building prototype proofs before talking about the idea, saves time and money 
Even that simple project isn’t as cut and dry and it seems with training required for the smaller to bring in its misses.
All in all it runs ok and pretty easy do though, technically speaking. Just takes a bit time wise.
I’m not sure the iterative processing and the reasoning chain that results would be as good as an o1 reasoning model tbh.
Why not just train that?