AI

• No model weights are in the APK.
• Private AI: connect an AI host you own on your network. Every resolved address must be private, and the request goes to the address that was checked.
• On this phone: download an allowlisted LiteRT-LM model (pinned Hugging Face commit and SHA-256) after consent; inference then runs on the phone. The allowlist fails closed if any entry is invalid.
• On-device generation has not yet been recorded on a phone.
• The AI advisor never signs transactions.