How every number is built.

Starts at 100. HRV against your personal band adds or subtracts the most; resting heart rate above baseline subtracts; sleep is penalised below about 7 h and above about 8.5 h; raised respiration and wrist temperature apply small illness penalties.

80+ Prime · 60–79 Good · 40–59 Moderate · below 40 Low.

HRV against your personal band is the primary signal. Form (TSB) and the illness panel can only downgrade a hard day toward rest; they never upgrade it. Easy days are never gated. With no HRV this morning, or no personal band yet, the call is unknown and the app says nothing rather than guess, unless an illness signal or very low form is real evidence on its own.

Only hard days are gated; the underlying HRV, TSB and illness numbers are always shown.

Combines your CTL (fitness) slope, ACWR (acute:chronic load) and Foster monotony. A strain or spike signal overrides a “productive” read (priority-ordered).

Detraining · Maintaining · Productive · Overreaching · Strained · Mixed.

An honest reserve from training load (TSB), HRV trend, resting-heart-rate drift and sleep debt, with each component shown. It refuses to fake a minute-by-minute battery.

70+ Fresh · 50–69 Balanced · 30–49 Accumulating · below 30 Depleted.

Every heart-rate sample recorded during the session is placed in your personal zones (from your LTHR when set, else your max HR); a sample stands for the time until the next one, capped at 60 s so a dropout can’t smear one reading across minutes. Session strain uses the same Banister TRIMP weighting and 0–21 scale as Day Strain, so the two compare directly. The intensity label is descriptive: easy (80% or more of minutes in Z1–2), threshold (35% or more in Z3), hard (20% or more in Z4–5) or mixed. It does not estimate a “training effect”. One-minute recovery is shown only when your watch measured it; it is never computed from an unstructured stop.

Well-trained endurance athletes do about 80% of sessions easy and 20% genuinely hard. A one-minute heart-rate drop of 12 bpm or less was the abnormal cut-off in a large exercise-test cohort; wrist readings after an unstructured stop run lower, so compare against your own sessions.

Three components, weighted and renormalised over whatever was measured: regularity (the Sleep Regularity Index below) 40%, duration 35%, efficiency 25%. Duration scores 100 inside 7–9 h (7–8 h from age 65), losing 30 points per hour short and 12 per hour over, floored at 40. Efficiency scores 100 at 90% or better, 80 at 85%, and 0 at 65% or below. Regularity follows the UK Biobank distribution: an SRI of 87 scores 100, the median of 81 scores 78, and 60 scores 0. A score needs at least two components; duration alone is not a score, and a missing component is shown as missing, never filled in.

85+ Excellent · 70–84 Good · 55–69 Fair · below 55 Poor.

Sleep-stage percentages: consensus supports continuity measures (efficiency, wakings, time to fall asleep) as quality indicators, with little agreement on architecture. Wake after sleep onset isn’t scored separately, because with only asleep and awake data it is the same two numbers as efficiency. A night whose source reports no awake time at all isn’t treated as flawless. Long nights are penalised more gently than short ones: the higher mortality of habitual long sleep is confounded by illness, and a single long night is usually catch-up.

The Sleep Regularity Index: the probability that you’re in the same state, asleep or awake, at the same minute on two consecutive days, scaled 0–100. Spline computes it minute by minute from the sleep your devices recorded, as the study did: waking in the night counts as awake, and naps count as sleep. Only nights exactly one calendar day apart are paired, so a missing night never compares Monday with Wednesday. It uses the latest 13 pairs within 28 days and needs at least 5.

85+ Excellent · 75–84 Good · 65–74 Fair · below 65 work on consistency. In the study, the median was 81; 85 sat near the 70th percentile and 65 near the 10th.

Spline borrows sleep staging from your device rather than owning it, so it flags low-confidence nights (phone-only detection, very short or fragmented sleep) instead of presenting stage percentages as fact. Nothing in the sleep score depends on staging, so the score stays valid on those nights.

Gross 7-day shortfall against a 7.5 h/night need. It reports gross shortfall rather than netting long nights one-for-one: a weekend lie-in only partially repays a weekday deficit.

Your wake time stays where it usually is: the median of your last 14 nights, rounded to five minutes and never moved, because a lie-in trades the bigger lever (regularity) for the smaller one. From it the plan counts back the sleep you need: 7.5 h, plus up to an hour while you are repaying debt, spread over several nights, divided by your own sleep efficiency, plus 15 minutes to fall asleep. Wind-down starts 45 minutes before lights-out and caffeine stops 6 hours before it. One night never moves your bedtime more than 45 minutes earlier than usual, and the plan needs five recent nights before it says anything. The same plan appears on the widgets and your watch in the evening.

A planning rule built on the evidence below, not a measurement; individual need spans roughly 7 to 9 hours.

Alcohol: drinks logged in Apple Health from noon yesterday to 6:00 today, at 14 g of pure alcohol each, divided by your body weight and placed in a dose band. Without a body weight, no band is claimed, because the bands are per kilogram. Caffeine: milligrams logged within six hours of your usual bedtime (the median of your last 14 nights). 200 mg or more is flagged; smaller amounts are only marked “watch”, because smaller doses weren’t tested.

Alcohol up to 0.25 g/kg low, 0.25–0.75 moderate, above 0.75 high. Overnight recovery fell by 9.3%, 24.0% and 39.2% across those bands in 4,098 within-person comparisons.

Readiness stays measured physiology. Mixing a self-report into a measured score is the black-box move Spline exists to avoid.

Candidate signals are ranked by how far each sits outside your personal range. The card cannot be shown without the measurement that triggered it: the reason is a required part of it, not optional text. It stays silent unless a usable HRV band and a readiness score exist, because “nothing is out of range” would otherwise only mean “nothing was measured”. There is deliberately no all-clear card: on a normal day, absence is the signal.

Asked only when a signal has persisted for three or more consecutive days, never after a one-day spike. Your answer becomes a journal tag on that day, which the pattern engine can later test. One question at a time, strongest first; answering or dismissing hides it for 14 days. It all happens on your iPhone.

Every habit and tag is compared against the next morning’s HRV, resting heart rate and sleep, never the same morning. A pattern is reported only if the effect is at least small-to-moderate (|Hedges’ g| of 0.3 or more), its bootstrap confidence interval excludes zero, each side has at least 10 nights, and it survives Benjamini–Hochberg false-discovery correction across every test run. Findings are worded as associations, never causes, with the option to test one properly as an N-of-1 experiment.

Daytime stress is derived from each day’s continuous heart rate and stored once a day from then on. The history reports the days above your own median, and whether the newest third of the window differs from the oldest third (thirds, not halves, so a fortnight-long shift isn’t averaged away). With fewer than 14 days it says nothing.

Each factor (VO₂max, resting heart rate, HRV, sleep, sleep consistency, body fat and activity) adds or subtracts years against a sex-specific healthy anchor. The factors sum to your offset, and each shows its own ± years. It needs at least three measured markers; a marker with no data is left out rather than counted as zero.

An estimate with wide individual uncertainty, labelled as one. The dial’s scale is honest: your calendar age sits at the top, with a ±15-year window either side.

The Levine PhenoAge formula: a weighted combination of albumin, creatinine, glucose, CRP, lymphocyte %, MCV, RDW, alkaline phosphatase, white-cell count and age, mapped through a Gompertz mortality model. All unit conversions are pinned by tests.

Lower than your chronological age is favourable. A risk surrogate with wide individual uncertainty, not a diagnosis.

Reference Change Value from each marker’s within-subject and analytical variation (asymmetric for right-skewed markers like CRP and triglycerides). Readings in different units or assays are refused rather than wrongly compared. Retest cadence follows the same biological-variation data.

Strong · Moderate · Emerging · Clinical

It never poses as medical advice: “Clinical” means bring the numbers to your doctor. Lp(a), for example, is framed as genetically fixed: the actionable response is to drive every other atherogenic particle low, not to chase a number lifestyle barely changes.

Your waist from Apple Health or a manual entry (which is written back to Health), divided by your height.

Below 0.5 healthy · 0.5–0.6 increased · 0.6 and above high.

Averages your recent home readings (a 7-day morning and evening average is the clinically valid number; a single reading is not a diagnosis). Category uses the ESC 2024 European scheme on home values (Non-elevated below 120/70, Elevated 120–134/70–84, Hypertension 135/85 and above), assigned by the higher of systolic and diastolic. ACC/AHA 2017 (US, home threshold 130/80) is shown as a comparison.

Mean arterial pressure; pulse pressure (systolic minus diastolic), flagged above about 60 mmHg as a large-artery-stiffness and independent cardiovascular-risk marker; and short-term variability.

Home target below 135/85. Risk is continuous: each 20/10 mmHg lower roughly halves vascular mortality, down to 115/75 mmHg.

A single-lead flag (sinus rhythm, possible atrial fibrillation, or inconclusive) is a screening signal, never a diagnosis. A possible-AFib result means: have a clinician confirm it with a medical 12-lead ECG; never start or stop medication based on the device. Inconclusive results are shown as such, not hidden.

Apple’s on-device model, or a built-in parser where Apple Intelligence isn’t available, only reads food names and amounts from what you wrote. Nutrition comes from food databases: Germany’s BLS, France’s CIQUAL, Denmark’s Frida, the USDA’s FNDDS (foods as eaten, with measured portion weights), and Open Food Facts for packaged products. Every portion shows where its weight came from: your own earlier correction, the food’s own portion, the label serving, a similar USDA food, or a density-aware estimate. Your corrections are remembered on your iPhone, where you can review or forget them.

A nutrition-label photo is the only place a number comes from an image, so it is cross-checked against the 4/4/9 kcal Atwater rule before you log it.

Your question is routed to a topic (readiness, sleep, training, HRV, a data lookup and so on), first by whole-word keywords in English and German, and only if that finds nothing, by Apple’s on-device model. The model may choose a topic; it never writes the answer. The app then composes the answer from the same values your cards show and the engines’ own verdicts (the gate, tonight’s plan, the strain target), so the coach and the cards cannot disagree. Missing data is stated as missing. A data question such as “when was my HRV lowest last month?” becomes a query over the same data the charts draw, and the interpreted query is shown above the answer so a misread question is visible.

Levers are drawn from a fixed, cited list and phrased as things to try one at a time, not prescriptions.