Tech & reproducibility
Everything here was run, not asserted.
Three levels of reproducibility, with exact commands and expected outputs in the repo. No remote API calls for inference — all OCR and MedPsy run locally.
1 · Data & grounding
Any machine, no models. npm run verify runs a 38-check suite over name resolution, DDInter lookup, the abstain logic, and the audit + resource-log writers — all deterministic.
2 · Real engine
On Linux/WSL2 with a MedPsy GGUF: live OCR reads a label → resolves → DDInter Major → a streamed explanation, plus abstain and no-fabrication checks on real inference.
3 · P2P mesh
An anchor provider + a consumer delegate a completion over the DHT and the answer arrives from the provider — proven across two physical machines.
The stack
Built end-to-end on QVAC
App
React Native via Expo (SDK 54 / RN 0.81), expo-camera, expo-sqlite for the interaction DB, device keystore for the shelf. Hermes JS engine.
Inference
Tether @qvac/sdk for all inference — OCR and MedPsy completion — running on-device via a Bare worklet, with GPU backends on Android (Vulkan/OpenCL). P2P delegation over Holepunch.
Data
DDInter 2.0 — ~160k drug-drug interaction pairs with severity, bundled offline as a generated SQLite database. Generic-name matching with a synonym table.
Evidence
A JSONL audit log of the grounded chain, a CSV resource log of on-device cost, and a packet capture proving zero outbound inference traffic offline.
Verified facts
What's been proven
- Grounded pipeline + abstain + no-fabrication + audit + resource log: 38/38 deterministic checks pass.
- Real engine run-validated on
@qvac/sdk@0.12.2: aspirin label → acetylsalicylic acid → DDInter Major, cited; paracetamol → abstain. - Cross-device mesh: a 4B-anchor delegation, two physical machines over the DHT, answer from the provider (local fallback disabled).
- Solo tier does zero network I/O during a scan after the one-time OCR-model fetch.
docs/REPRODUCIBILITY.md in the repository.The repository is made public, Apache-2.0, at submission.