A local AI lab that runs experiments on your life — and tells you, with proof, what actually makes it better.
Recovery. HRV. Steps. People grind a score up while the thing it was supposed to stand for — a life that feels good to be inside — quietly slips. That's Goodhart's law wearing a fitness band: when a measure becomes the target, it stops measuring anything that matters.
Longevity, recovery, sleep — all of it is a means. The end is whether your days are worth having. This system keeps that end in the frame on purpose, because it's the one thing a tracker structurally can't: a wearable doesn't know what your good life looks like. You do — so you name it, and everything else is held accountable to it.
The top metric isn't a body score.
Is my life actually better?
When the body score climbs but life doesn't, that isn't a win to celebrate. It's metric tyranny — and the system flags it instead of congratulating you.
Built for people who've tried enough supplements, trackers, and protocols to know the hard question isn't "what's optimal?" but "what's optimal for me, in the life I actually live?"
Correlations are guesses — "you sleep worse when you drink" can't tell if the drink did it or a hard day caused both. So the harness borrows the method clinical medicine uses to decide if something works for one specific person: the n-of-1 trial, pre-registered, controlled, judged against a line drawn before the data arrives.
A specific causal claim.
Hold everything else. The rule everyone breaks.
Written before the data, so you can't fool yourself.
Merge it, or revert and log why. A washout confirms.
And every claim the agent makes carries its epistemic weight — observation split from interpretation, each line cited to its source:
“Creatine moved nothing for you in three weeks — stop paying for it.” — the kind of sentence no wearable will ever say
Not a medical device — and not invented. The discipline it enforces is the one clinical science uses to decide whether something works for one specific person.
Single-subject design from personalised medicine: pre-registered, one variable, washout, verdict.
Anchored on a patient-reported outcome (PRO), not a proxy like HRV. Validated tradition (WHO-5).
Every claim labelled fact / inference / assumption and cited — observation split from interpretation.
Logging in the moment is ecological momentary assessment — it beats recalled, reconstructed diaries.
Full clinician brief, limitations, and the rest of the references — in METHODOLOGY.md ↗
Pick a change. The harness holds one variable, measures against a line set in advance, and returns a verdict — merge, or honestly, revert.
Text, voice, or a photo — selfie, food, stool auto-sorted on-device. Several a day. Zero forms.
Hybrid RAG retrieves your real history; the agent finds patterns you can’t see and writes them down with sources.
A hypothesis turns into an n-of-1; a verdict turns into a standing rule — knowledge that compounds.
Talk, type, or snap. Selfie, food, and stool photos route themselves into the right diary with on-device vision — no captions, no forms, nothing uploaded.
Morning brief, a question or two across life domains, an evening recap. Mute it anytime.
The agent reads your week and scores it against your north-star.
Run, extend, or log a lapse. One active at a time — truly one variable, so the effect is attributable.
Tell it what’s already standing — 15k steps, omega for years — and it won’t re-discover them.
Watch quality of life move over seasons as data compounds.
Your whole life is a folder of plain Markdown you own — a layered pyramid where evidence flows up and nothing is asserted without a trail.
A wearable is a brilliant sensor — for physiology. But the drivers of how a day actually feels live mostly outside its straps. Flip it:
Toggle what a week held. Watch what actually moves "is life better?" — and what just looks like progress.
Capture is free — everything on-device. The only spend is the agent's analysis. Move the dials:
Capture is on-device: voice→text, photo sorting, and search all run locally, and your files live on your disk (photos and sensitive diaries are never pushed to git). Two things do leave, by design: messages transit Telegram, and deep analysis uses Claude when you ask the agent. Nothing is sold or posted.
The capture loop is free — bot, on-device models, and storage cost nothing. The only spend is the weekly analysis: included if you run the agent on a Claude subscription, or cents per session on a raw API key. A photo costs well under a cent.
No. The agent describes and compares; it never diagnoses. Any worrying sign — from a stool photo or anything else — resolves to a single recommendation: see a specialist.
No. Body data arrives on its own; the rest is whatever you feel like sharing. Skip days freely — it's a tool for your life, not another job.
People who've tried enough supplements, trackers, and protocols to want what works for them, in the life they actually live. Not for streaks, badges, or a prettier sleep chart.
Open-source, local-first, built to generalise to any self-quantifier.