The new architecture is designed to work without an account, tracker or sending content to an AI provider.
The browser loads a public versioned JSON bundle of concepts, questions and rules from Kerkito. Your text analysis, answer-contract compilation and structural validation run client-side; raw content is not required by the knowledge API.
By default, the project remains in session memory. Local persistence is triggered voluntarily by the user. Server synchronization is not required for core operation.
Kerkito does not call OpenAI, Anthropic, Google or another LLM to interpret content. A prompt can be exported later if the user chooses to work with an external tool.
No advertising pixel, tag manager or third-party analytics is loaded by the application. Hosting technical logs may still contain the usual network metadata required to operate the service.
Interpretation confirmations and rejections are stored only after an explicit user action. They remain on-device, can be cleared and are not used to build a remote profile.
Voice playback and dictation are optional. Kerkito detects what the browser exposes. When local recognition is available it is preferred; otherwise the interface states that the browser provider may process audio. Kerkito does not receive dictation audio.
Remote synchronization is not required for Kerkito 8.3. If it becomes useful later, it must be designed separately with browser-side encryption and documented limits before making any claim about what the server can or cannot know.
