Your question—'who watches the watchers'—is the exact focal point of the JTP framework. In many current systems, the feedback loop between the Supervisor and the Base model is 'silent' and internal. My concern is that even when the Supervisor works perfectly, the human user remains in the dark about where the system corrected itself.
Instead of the Supervisor's feedback being a background process, it could be surfaced to the user as a 'trace' or a 'seam'—allowing the user to actually perceive the internal deliberation. This turns the human from a passive recipient into the final, informed 'watcher.'
I’d be honored to discuss how these JTP principles might serve as a transparency layer for your work. I’ll be keeping a close eye on your repository!
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).
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.