Philosophy

Seven principles for AI
that earns the right to live near a life.

Most wellbeing apps are habit machines wearing a meditation costume. We chose, early, not to build one. The seven principles below describe the design discipline that follows from that choice — the same discipline that shapes every ritual, model, study and reply we put out into the world.

I

Restraint over reach

We will not optimise for daily sessions. We will optimise for restful weeks. The first metric is composite — subjective wellbeing, sleep quality, sense of agency — measured at week's end. Daily engagement is, at most, a diagnostic.

II

Quiet by default

Notifications are a withdrawal from a finite attentional account. Our model is trained to send fewer, never more, and to withhold itself when the user's body signals recovery rather than stimulation. The default is silence; speech is earned.

III

On-device for the intimate

Journal entries, biometrics and reflective text never leave the device. We run distilled, small models locally — at acceptable fluency for the work — because the alternative exposes the most intimate dataset a person produces to a class of risks we cannot rule out.

IV

Personalisation without surveillance

Our personalisation operates on weekly summaries and locally derived embeddings. We do not need your day-by-day timeline, and we deliberately have not built the infrastructure to receive it. What we cannot collect, we cannot betray.

V

Honest about trade-offs

On-device models are not as fluent as the largest cloud models. We will not pretend otherwise. We have evidence that for reflective journaling the trade is worth it; we publish the trade openly, with the caveats, and let users decide.

VI

Slow publishing

Our cohort report took fourteen months. Our manifesto took nine drafts. We would rather publish less and be trusted than publish often and be ignored. The journal is updated when there is something worth saying, not when a calendar demands it.

VII

Designed to be uninstalled

Our internal success metric is the share of users who, six months in, report that they need the app less. We optimise for that number. We tell users, in writing, when we think they are ready to step away. The grove ends. That is the point.

A technology that respects you
is one that knows when to stop talking.

— A founding note, drafted in the Kings Cross studio, 2025

Read on

The principles, applied

Each ritual, each model card, each published number connects to one of these seven principles. Start where you are curious.

The Quiet AI essay → Cohort report → On-device note →