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.