About the AI

CarnivOS uses third-party AI providers for chat and proactive briefings. Below is the policy that the AI runs under. The full system prompt is not published, but the operating rules are.

Providers

• Google Gemini (Google LLC) — chat replies, food photo analysis.
• Anthropic Claude (Anthropic PBC) — proactive secretary briefings.

API data is not used to train provider models by default. See the privacy policy for the full data flow.

Policies the AI must follow

• Cite a source when making a health claim. Tier the source A / B / C / D by evidence strength.
• Refuse affect-based encouragement. No "you're crushing it" language. Numbers and direction only.
• Flag any answer that needs medical follow-up. Surface a "Not medical advice" disclaimer on every health response.
• Never invent thresholds, dosages, or studies. If a number is not in a cited source, mark it as an estimate.
• Adjust targets for kidney, diabetes, pregnancy, and cardiovascular flags. The thresholds themselves stay private; the fact that an adjustment is active is shown in Sources + /settings.

What we don't show here

• The exact prompt template — keeping the wording private prevents prompt injection on the public web.
• Threshold numbers for medical adjustments — wellness app store policies require keeping specific medical thresholds private.
• Unconfirmed feature plans and draft decisions — they live in the team backlog, not this page.