OrigenStudio wants OpenAI to never see raw client data

On paper it's just documentation - a diagram of how the fork is wired together. But the interesting part is what it telegraphs about where the team is heading. Alongside the known hosting stack, the diagram sketches a privacy layer that hasn't been built yet: a step that scrubs personally identifiable information from requests using a small AI model running on the team's own hardware, so that whatever reaches OpenAI is already stripped of raw client detail.

Nothing here has shipped. The design is explicitly marked as planned, and the team has paused at an internal approval gate. So treat it as intent, not capability. Still, for anyone weighing whether a legal-AI tool can keep confidential matter away from a third-party model, a self-hosted scrubbing step in front of the LLM is a pattern worth watching.