This is exciting, I hope you manage to get traction for the idea! I currently have rely on a sort of supervisor LLM to check and detect if we're drifting, or o

This is exciting, I hope you manage to get traction for the idea! I currently have rely on a sort of supervisor LLM to check and detect if we're drifting, or overcomplicating or similar (https://github.com/open-horizon-labs/superego https://github.com/open-horizon-labs/superego). While I still to figure out who watches the watchers, they're are pretty reliable given the constrained mandate they have, and the base model actually (usually) pays attention to the feedback.