Repository navigation
[Feature] Evidence-linked AI observations and qualified local providers #46
Description
Outcome
Finish AI as an optional evidence narrator inside Today and Trends—not as a blank chatbot—and add qualified on-device provider choices without weakening the deterministic product.
Dependencies
- Deterministic observations and evidence: [Feature] Deterministic metabolic observations and evidence model #37
- AI evidence/safety/provider contract: [Feature] AI evidence, safety, and provider contract hardening #39
- Unified Today/timeline drill-down: [Feature] Unified body timeline and compact Today experience #44
- Source-aware data and context features: [Feature] Source-aware health data spine and import identity #36, [Feature] Fast journal and meal/activity logging #41, [Feature] Apple Health context import #43, [Feature] Android Health Connect context import #42
Product surface
- “What stands out” card with at most 1–3 observations.
- Daily and weekly user-triggered summaries first.
- Each statement has evidence chips and opens the supporting trace/window.
- Show model, generation time, prompt/evidence version, data coverage, and local/remote status.
- Regenerate, hide, delete, and feedback actions.
- Deterministic text remains available with AI disabled.
Provider delivery
- Swift bridge for Apple
SystemLanguageModelon supported iOS versions/devices, with explicit availability reasons and no deployment-floor increase. - Pinned llama.cpp/
llama_cpp_dartspike for qualified iOS/Android devices. - Curated GGUF catalog only: immutable revision, file hash, license, size, context, device tier, and benchmark evidence.
- BYO cloud remains Advanced with endpoint/data preview.
- Never silently switch local work to a remote provider.
Resource and safety controls
- Unload local models on memory pressure/background and stop at serious thermal state.
- Bound context/output, disk, memory, and model concurrency.
- Test BLE session, live notifications, background/resume, battery, and model load/unload coexistence.
- Model output may phrase verified observations but cannot invent calculations, trigger alerts, alter records, issue sensor commands, or provide diagnosis/treatment/dosing/emergency advice.
Parallelization
Blocked for product integration. Native provider spikes may begin after #39 stabilizes provider capabilities. The visible card waits for #37 and the body-timeline drill-down.
Acceptance criteria
- Every factual statement links to valid deterministic evidence.
- AI-off mode retains full analytics and deterministic observation value.
- Local provider availability and unavailability reasons are honest.
- No local-to-cloud fallback occurs without an explicit user action.
- Model provenance, license, hash, disk size, resource state, and data recipient are visible.
- Stale insights invalidate when underlying evidence changes.
- Delete/export removes all outputs, conversation state, models, and keys as selected.
- Prompt/model/OS evaluation matrix and unsafe-input suite pass.
- Physical devices demonstrate bounded memory, thermal, battery, BLE, and background behavior.
- R2 independent review and rollback plan are complete.
Out of scope
- Open-ended chat until evidence retrieval, citations, retention, and prompt-injection controls are proven.
- Medical decisions or autonomous actions.
Activity
shroominic commentedon Sep 1, 2026
Superseded by #99, recreated by the repository owner under
shroominic.