The architecture stays. The evidence gets richer.
Dates are intentions, not commitments. The shape is more honest than the calendar. Each phase below is something the twin gets noticeably better at, from a user's point of view.
01The brain works end to end.
ShippedSix agents on a typed state graph. The Auditor closes the learning loop. The Forecaster writes predictions and the calibration math updates after each one comes due.
02The iPhone client and the App Store launch.
Wiring upThe thin iOS client talks to your container over a WebSocket and REST. Subscription via the App Store. Apple Sign-In runs at the purchase moment, only then.
03Time-series priors.
DesignedSmall numerical models feed structured priors into the Forecaster prompts. Better hints for sleep and HRV without changing the agent architecture.
04A Dardar of you.
On the mapAfter hundreds of confirmed predictions, a small open-source model can be fine-tuned on your history alone. Runs inside your container. Strictly per user.
05Multimodal evidence.
On the mapPhotos for meals, skin, and eyes. Voice tone as a soft mood signal. ECG morphology. All feed the Perceiver and enrich the Forecaster.
06Privacy-preserving population insights.
On the mapFederated learning lets twins contribute anonymized calibration data without sharing raw signals. Population priors become a soft hint, never a fact about you.
07Offline operation.
On the mapThe same agents bundled inside the iOS app for users who want zero-network operation. Same Python code, different runtime.
We will not promise a feature here until the work is real. Anything not on this page is not on the roadmap.