DailyVox is built around a single hard invariant we call the Never-Leaves Guarantee: the user's journal content — audio, transcript, features, embeddings, model weights, and adapter parameters — never leaves the device, under any operating condition. The phone captures, transcribes, analyzes, stores, infers, trains, and speaks. The work below — a preprint, an open-source package, and two research reports — makes the architectural case.
Four primary artifacts. Each one stands alone and cross-references the others.
The paper introduces the Never-Leaves Guarantee — an architecturally enforced invariant that journal content never traverses the network for any purpose, current or future — and shows how it is maintained across every layer: unconditional on-device transcription with a knowledge-graph-informed correction loop, four online personality estimators with bounded memory, a typed sentiment-weighted knowledge graph with Levenshtein-based identity reconciliation, and a roadmap where Foundation-Model inference, LoRA fine-tuning, RAG, and acoustic-feature extraction all execute on the phone.
Link available on publicationClean, dependency-free reference implementation of the personality-modeling engine described in the paper. MIT-licensed, zero third-party dependencies, multi-platform (iOS, macOS, visionOS, tvOS, graceful Linux fallback), Sendable-annotated for Swift 6, 7 passing tests, CI on macOS + Linux. The shipping app currently runs an in-tree copy of the same algorithms and will consolidate on this package in a near-term release.
View on GitHubAnalyst-grade snapshot of the category. How AI journaling apps collect, transmit, and process user data in 2026; the four-tier architectural classification; the economics shift from cloud to on-device; five concrete 2026–2027 forecasts.
Read the report40 journal and AI-journal apps classified into four tiers by architecture, not policy. Includes observation methodology, vendor-response process, and citation-ready BibTeX. The reference the next journalist roundup will quote.
Browse the auditAll four artifacts are citation-ready. Copy the block that matches what you're referencing.
If you're writing about on-device AI, privacy-preserving personal informatics, voice journaling, or Apple's Neural Engine / Foundation Models ecosystem — the four artifacts above are free to excerpt, translate, and republish with attribution (CC BY 4.0 where applicable; MIT for code).
Raw audit data (40 apps × 12 attributes), source LaTeX, benchmark methodology notes, or a quote for a piece you're working on: email intrepidkarthi@gmail.com. Responses within 48 hours.
Preferred author credit: Karthikeyan NG, independent researcher & author of DailyVox.