The Science · Baku
The science

Published models,
run on your own data.

Baku does not invent physiology. It implements models that already exist in the literature — and the interesting engineering is in running them correctly on one person’s data.

Sleep pressure

Debt does not clear in one lie-in.

Sleep debt follows the Van Dongen (2003) framework: a chronological running balance over 14 nights. A short night adds to it, capped at two hours — and a long night repays only half of its surplus, because recovery from accumulated restriction is partial.

HOURS SLEPT · NEED 8.0H RUNNING DEBT 14.8H BY NIGHT 14 a 9.6h night, and the debt barely moves

A surplus night repays half its surplus — debt does not clear in one lie-in

Need, then deficitYour need is learned from your own history, not fixed at eight hours for everybody.
Capped nightlyOne catastrophic night cannot dominate a fortnight of ordinary ones.
Partial repaymentWeekend recovery sleep genuinely helps and genuinely does not reset you. The model says both.
Training load

Fatigue clears before fitness does.

The Banister (1991) fitness–fatigue model, with the two time constants that make it work: fitness decays with τ = 45 days, fatigue with τ = 15 days. Form is the difference between them, which is why you feel worst mid-block and best a fortnight after it.

TRAINING BLOCK ─ FITNESS τ 45d ─ FATIGUE τ 15d 38 DAYS

Fatigue clears first, which is why form rises after the block and not during it

Two decay ratesOne number cannot express “fitter but tired”. Two exponentials can.
Acute against chronicLoad balance compares your last 7 days with your last 28 — the ratio that flags a spike.
Anchored to your maximumTraining impulse is computed from heart-rate reserve against your own measured maximum, not 220 minus your age.
Baselines

Your normal, not everyone’s.

Every reading is judged against your own recent range rather than a chart of what is normal for everyone. A pulse of 66 is unremarkable for an adult and a warning sign for someone who wakes at 54 — and a jump like that is one of the earliest signals there is that something is coming.

THIS MORNING · 66 BPM YOUR MORNINGS ADULTS A COMPLETELY ORDINARY PULSE FOR AN ADULT 12 BPM ABOVE EVERY MORNING YOU HAVE RECORDED

The same number, read two ways. Only the second one would make you rest.

Rolling, not fixedBaselines move with you. Get fitter and the bar moves, which is the point.
Compared to you, not to a ruleThe range Baku judges you against is built from your own nights, so no number fitted to somebody else gets to decide how your day went.
Built on your own weekReadiness rests on seven nights of your sleep before it reports a score.
Independence

Every signal counts once.

Following the Altini critique: HRV already encodes much of sleep quality, so counting both at full weight penalises one bad night twice. Where HRV is estimated rather than measured — and an estimate derived from resting heart rate is not independent of it — that term is dropped and the rest reweighted.

An estimate is not a measurementA watch that reports no HRV gets its weight redistributed, not a plausible-looking stand-in.
Never a silent zeroA component with no input contributes nothing. It never contributes zero, which would read as “bad”.
Stated on screenWhen a score rests on fewer components, the app says so rather than presenting it identically.
Devices

Two devices, one scale.

A watch and a ring measure the same night and disagree — not by noise, but by a systematic offset in how each scores stages. One device is pinned as the reference, and Baku fits the offset at four points across the distribution, not just the middle. Because the correction is multiplicative, every night’s standing against every other night on the same device is left exactly as it was — only the level moves.

× 1.25 RING · 57 WATCH · 71 SLEEP SCORE DASHED = RING, CORRECTED

Same nights, two devices, one continuous record. Measured on a real switch: the ring’s scores land on the watch’s scale, so the history reads as one person rather than two.

Per-device bandsYour normal range is worked out separately for each device you wear, so changing hardware is not mistaken for a change in you.
Composite corrected, components notThe composite score is calibrated; raw stage percentages are deliberately left uncalibrated and simply not trended across the seam.
Both devices at onceWear a watch and a ring together and each night is attributed to the source that actually measured it.
Uncertainty

The range is part of the answer.

Biological age is reported with a plausible interval, and that interval is wider when fewer markers are present. It narrows as data arrives. An estimate that visibly sharpens is the method working.

Absent markers widen itWhere a driver is missing, the range grows to cover where it plausibly sits.
Coverage is trackedEach input carries how many days stand behind it, and that feeds the width.
Width tracks evidenceA two-day-old account gets a genuinely wide range, and every day of data narrows it.
Limits

What Baku does not claim.

This matters more than anything above it.

Consumer wearables are not medical devices, and Baku is not one either. It cannot diagnose anything, and nothing here is medical advice — if something concerns you, that is a conversation for a clinician. Estimated values are estimates: VO₂max is inferred from heart-rate response, not measured in a lab, and sleep staging from a wrist or a finger is an approximation of what a polysomnogram would score. What consumer sensors are genuinely good at is relative change in one person over time, and that is what Baku is built to read. Whether acting on that improves any individual’s health is not something we can promise, and we will not.