Context became a private learning loop
Available in private Mac betaSideQuesto now turns privacy-safe work outcomes into selective lessons that strengthen one shared narration policy without making human ratings the engine of improvement.
- Verified outcomes are evaluated against evidence, timing, and audience fit, then distilled into categorical lessons that contain no prompt, code, path, log, transcript, or narration text.
- Relevant lessons are selected for the current task instead of replaying complete histories or treating every past event as equally useful.
- Dex keeps a compact expert contract while Pip can recognize a topic-level foundation question and explain it without permanently labeling the person.
- Human feedback remains an optional calibration signal for subjective qualities; it is no longer the only route to measurable learning evidence.
- One compiled narration policy remains active. Derived lessons are encrypted, versioned, reversible, and unable to bypass privacy, attention, or promotion gates.
- Mastery Lab now separates raw benchmark quality from verified evidence scope with a layered radar profile, visible future-suite rings, axis scores, safety status, and weighted L100 target rails.
- The release adds contextual-learning, audience-boundary, encrypted-persistence, selective-retrieval, and policy-regression coverage to the complete Swift suite.
SideQuesto improves its local context policy; it does not train or modify the underlying foundation model.
