Comment by calmbonsai

This is intriguing. I've been working on gestalt measurements of LLM 'skew' and thought this sort of scenario might be possible.

Combine that workflow with homomorphic encryption and you've got a reasonable privacy moat.