What is shipped,
and what is not.

Fifteen releases in five months. The interesting column is the right-hand one: most of these turns were forced by a measurement rather than chosen from a backlog.

DailyVox v1.11 is on the App Store, free, for iPhone and iPad running iOS 17 or later. Voice journaling with fully on-device transcription, the Digital Twin, the constellation, Ask Your Twin, Insights, encrypted export, biometric lock, widgets, Live Activities, Siri Shortcuts, read-aloud, live transcription in the Dynamic Island, a Control Centre control, share cards, and English, Spanish, French, German and Italian.

The claim that matters is the one you can test in ten seconds: turn on airplane mode and record an entry. Everything still works, because there is no server to be cut off from.

Not released. Not on the Play Store. No date announced. Saying otherwise would be the easiest sentence on this page to write and the one you would most deserve to be annoyed about later.

What is true: a native Kotlin and Jetpack Compose app exists, it is feature-complete against the iPhone version in development builds, and it holds no internet permission of any kind. Not "we don't use the network" — the permission is absent from the manifest, so no version of that build could send a journal anywhere even if its code tried.

Two decisions are worth stating because they cost something:

Android's automatic cloud backup is disabled. Left at its default it would have copied the journal database and every audio recording to the user's Google Drive — performed by the operating system, needing no permission from the app, and leaving every privacy claim technically intact while the diary sat on a server. Phone-to-phone transfer during device setup is still allowed, because it is local and because losing years of diary when you change phones is its own kind of harm.

There is no free-form chat with the Twin on Android. An on-device model small enough to ship to the phones this audience actually owns would guess, and a Twin that guesses about your own life is worse than one that stays quiet. Ask answers structured questions from real statistics and cites the entries it used.

The honest gap: the entity graph depends on Android's speech recogniser capitalising names, and that has not been verified across a broad range of physical devices. It is the single largest untested assumption in the port, and it is why there is no date on this section.

Neither platform goes straight to v2.0. There is a run of releases in between on both, and what gates most of them is a measurement rather than a feature — the version numbers move when a number does.

Body Twin. Sleep, heart-rate variability and activity as context the Twin can correlate against mood — HealthKit on iPhone, Health Connect on Android. Opt-in and read-only on both. The interesting question is not whether poor sleep makes a bad day; it is whether it does so for you, which is a question only your own entries can answer.

Personality depth. Validated Big Five scoring from journal narratives, with the Twin's replies conditioned on your measured profile and your actual phrasings rather than a generic voice.

A true digital self. After years of entries: a Twin that predicts your reactions, explains causality from your own past, and shows how you have changed. Not a clone. It knows your narrated self — the thoughts you never journal are invisible to it, and always will be.

A roadmap without these is marketing.

Voice cloning is off the roadmap. A candidate on-device model was evaluated and closed after a blind test: real versus synthetic samples were distinguished 8 times out of 8 (p = 0.0039). An earlier approval of that model had been based on it sounding better than the alternatives, which is not the same as sounding like a person.

A learned emotion head was dropped. A model trained on public emotion corpora failed its accuracy gate twice, and a ten-configuration sweep found no setting that passed. The reason is register: a model carries whatever prior its training data had, and diary writing does not sound like the internet. A lexicon has no corpus, so it has nothing to mis-transfer — which is why the lexicon approach transferred where the trained head did not.

Multi-platform was reversed. macOS and visionOS were on an early roadmap and are not coming. Every feature must work end to end on a phone alone; nothing may ever require a second computer.

DailyVox has a research programme attached to it, because a model built to be private cannot be watched in production — with no egress there is no dashboard, no A/B test and no confusion matrix, so it has to be evaluated on the device before it ships. That argument, the paper it produced, the evaluation harness, what the harness found against this app's own model, and five open problems anyone can pick up are all on the research page.

What belongs here is narrower: which of those lines is actually moving.

What is actually being worked on, and what is parked

This is a one-person project with a day job, so the research is filtered by a single question: does it ship as a feature? Work that produces only a paper competes with the product. Work that produces a release and a result compounds. Two lines pass that filter and two do not, and it is more useful to say which than to list everything as "planned".

The honest cost of this filter: the two parked lines are the two that would turn "it works for me" into a finding. Shipping features is the easier half, and it is the half being done.

"Digital Twin" is a claim, not a feature, and the only honest way to make it is to say in advance what would prove it false. Each rung is a question with a number attached. They are in order: a Twin that mishears you cannot read you, and one that cannot find the right memory cannot model you.

1. 01

   Does it hear you?

   measured

   On-device transcription, and the willingness to fail rather than reach for a network when the model is missing. This is the rung everything else stands on, and it is the one the app got wrong until v1.11.

2. 02

   Does it read how you felt?

   measured

   Against 1,459 human-labelled entries the bundled lexicon scores r = +0.663 with 87.1% sign accuracy, ahead of the platform's own +0.594 at 79.8%. Shuffled controls collapse to chance, which is the part that makes the number mean anything.

3. 03

   Does it know who is in your life?

   measured

   99.1% of 111 hand-annotated name spans found, at 61.6% precision. Every person name in that corpus is non-Anglo, and all of them were found — a learned model missed 2.3% of them, because it has a prior about what a name looks like and that prior is parochial.

4. 04

   Can it find the right memory?

   measured

   Semantic retrieval with a measured abstention threshold, so the Twin declines rather than guesses. An answer that cites the wrong entry is worse than no answer, because you cannot tell which kind you got.

5. 05

   Does it remember across years?

   in progress

   This is the line being worked on now. Retrieval over a week is not memory. The open work is long-horizon evaluation — the kind of benchmark that asks a question answerable only by connecting entries months apart, and scores whether the connection was made or confabulated. A journal's value compounds; the evaluation has to as well.

6. 06

   Does it model who you are, not just how you felt?

   in progress

   Mood is a reading of a day. Traits are a claim about a person, and they need validating against instruments built for the purpose rather than asserted from a paragraph. Until that is done the Twin reports a Twin Resolution score: how well it thinks it knows you, stated so you can disagree with it.

7. 07

   Does any of it hold for someone who isn't me?

   not yet

   Parked, and honestly so. Every number on this site comes from one author's diary. That is enough to build with and not enough to conclude from. A cohort — many journals, many people, the same measurements — is the gate between "it works for me" and a finding, and it is the one rung that cannot be climbed by writing more code.

Rungs 1 to 4 ship in the app today and the numbers above are reproducible from the source. Rungs 5 and 6 are where the work is. Rung 7 is the honest reason nothing here is described as proven.

Retention is the binding constraint, not acquisition. Install conversion is healthy; daily active use is not. That is a product problem, and no amount of new features fixes a journal people do not come back to.

Sync does not exist, and might not. There is no iCloud equivalent on Android, and any cross-device sync needs a network permission that would end the argument this product is built on. Manual encrypted export works and is portable between platforms — a backup written by the iPhone app can be opened by the Android app and the reverse — but it is a manual act, and manual acts do not happen.

The Twin has never been measured against other people. Every number quoted anywhere on this site comes from one author's diary. A cohort study is the only thing that would turn "it works for me" into a finding, and that is the point of the research programme.

Try the thing itself

The fastest way to judge any of this is to record one entry in airplane mode.

Download free on iPhone →

Android is in development and is not available yet. There is no waitlist and no email capture — when it ships, it will be on the Play Store.