Velisa refuses to guess.

A 57 ms HRV reading is not good or bad in the abstract. It is above or below your own recent middle. Every engine builds that baseline the same way — the median, not the mean, and the median absolute deviation, not the standard deviation. One three-hour night should not be allowed to redefine what is normal for you.

z  =  0.6745 × (value − median) / MAD

Velisa/Scoring/Baseline.swift:24 — 0.6745 rescales MAD so it is comparable to a standard deviation. isReliable is count >= 14: below fourteen days of history, z() returns nothing at all and every engine that depends on it declines to render a verdict.

Two rules govern the published tables underneath that. Thresholds are health-optimised, not population averages — anchoring to a population mean is exactly why other effective-age calculators tell the average adult they are biologically younger than they are, because the mean is dragged down by the unwell. And a missing interval is never invented: where a paper publishes only a point estimate, that contributor is recorded as carrying no disclosed error, and the output states how many contributors were in that position — so “no uncertainty shown” is never quietly confused with “no uncertainty exists”.

SourceVelisa/Scoring/References.swift, file header — “Nothing in this file is fitted, tuned or guessed.”

“The age of a person whose all-cause mortality risk matches yours, given your fitness and a short list of measured behaviours.” The spine is your fitness age: a published VO₂max reference table is inverted, so instead of asking what you should hit, it asks at what age the reference value equals what you actually have. That is a purely descriptive statement and needs no assumption about how fast anything declines.

The ledger is then added on top. For each behaviour with a published hazard ratio — sleep regularity, sleep duration, steps, moderate and vigorous activity, strength training, resting heart rate, body fat — the ratio converts to years through Spiegelhalter's effective-age transform. All-cause mortality rises about 10% per year of age after 25, so:

Δ years  =  ln(HR) ÷ ln(1.10)   ≈   10 × ln(HR)

    HR 1.22  →  +2.0 years
    HR 0.85  →  −1.6 years

Each Δ is shrunk by how well Velisa actually measured that exposure, capped at ±5 years per contributor and ±15 in total. A contributor you have no data for contributes exactly zero years and widens the ± — it is never filled in with a population average.

Spiegelhalter 2016 · WHOOP Healthspan 2025 · Kaminsky 2015 (FRIEND) · Nes 2011 (HUNT)

Four overnight markers — HRV (log-transformed, because raw HRV is right-skewed), resting heart rate, respiratory rate and sleep adequacy — each turned into a personal robust z-score against a 28-night baseline, then combined with equal weights.

composite Z  =  mean(z₁, z₂, z₃, z₄)

unit-weighted — no component is favoured

“The weights are equal, and that is a choice with a paper behind it rather than a shrug: unit-weighted models match or beat optimally-weighted ones when the criterion is noisy and the true weights are unknown (Dawes 1979). Polar and WHOOP publish contradictory orderings and no numbers.”

Shown inside the app, in the Recovery Score explainer · Velisa/Scoring/RecoveryScore.swift:838

The ± around it is the wider of two things: how far the number would move under Polar's stated ordering or WHOOP's stated ordering instead of unit weights, or the device's own published overnight measurement error. Whichever is larger is named as the reason for the band's width — a false sentence about why a number is uncertain is still a false sentence.

And four conditions under which it refuses to answer at all. Each one is a case where the arithmetic would still run and the answer would still be wrong, which is exactly when a health app is most dangerous.

• Two readings of the same night. HRV below baseline and resting heart rate also below baseline. Both the fatigue reading and the parasympathetic-saturation reading are live, and the literature does not sign that quadrant.
• Illness signals own today. A recovery score is answering the wrong question, and “you're at 22%” reads as “train less” when the correct statement is “three overnight signals moved together”.
• Your HRV baseline restarted. The source or convention changed inside the comparison window — Apple stores SDNN, Whoop reports RMSSD. The step is a device swap, not you.
• The weighting decides this number. If the sensitivity band spans 20 percentile points or more, the four components are shown instead of a single figure.

Below that last one the app says so, in these words: “A percentile computed from fewer is a number pretending to be one, so this shows the band and the four signals instead.”

Doherty 2025 · Bellenger 2016 · Lundstrom 2024 · Dawes 1979 · Dial 2025 · Polar Nightly Recharge 2019 · Plews & Buchheit 2013

“Energy” is not a measured quantity. It is four different real things wearing one word — metabolic energy, sleep pressure, autonomic arousal, and perceived fatigability — and only the first is genuinely conserved. There is no physiological reservoir that drains from 100 to 0 over a day.

So a published alertness curve carries the spine, and three ledgers sit beside it. They are never summed, with the curve or with each other:

lane 1   minutes above your resting range
lane 2   training load added today
lane 3   kilocalories in versus out
──────────────────────────────────────────
total    —  different quantities, different
            units, and no published factor
            converts between them

The heart-rate lane is labelled “time above your resting range”, not “stress”, on purpose: the classifier it would be imitating reads beat-to-beat HRV, and Velisa has heart-rate bins. Calling it the same thing would be the false-precision this metric exists to avoid.

The spine is the Three-Process Model, and checking it found an error. Velisa's own research memo prints the two homeostatic equations with their asymptotes swapped. As written, the wake process rises and the sleep process diverges. The memo's own definition of the recovery rate settles which way round they go, because that expression is the algebraic solution of one equation and not of the other:

g  =  ln((ha − 14.0) / (ha − 7.96)) / 8
   =  ln(0.3 / 6.34) / 8
   =  −0.38136

used:  S (t_aw)  =  la + (s_w − la)·e^(d·t_aw)
                   → decays toward la = 2.4

       S′(t_as)  =  ha − (ha − s_s)·e^(g·t_as)
                   → recovers toward ha = 14.3

Every constant is printed in Ingre et al. 2014 — 136 aircrew, 964 sleep logs. None is fitted here. The correction is not merely written down: the test suite re-derives the published 7.96 → 14.0-in-eight-hours anchor from the implementation, so it is asserted rather than asserted-about. Where a form is not published — Ingre's two brake constants are, the algebraic form that consumes them is not — that gap ships as a visible disclosure on the lane itself.

Åkerstedt & Folkard 1997 · Ingre 2014 · Enoka & Duchateau 2016 · Tanaka 2001 · Fuller 2020

Every metric carries its own minimumWindow string, and the type will not compile a metric without one. These are those strings, unedited.

Ten of eighteen. The remaining eight — volume load, estimated one-rep max, weekly sets per muscle, Velisa Trajectory, sleep debt, chronotype, trend weight and macro targets — carry the same field and the same rule.

Things the literature or the data genuinely does not support, listed on purpose rather than quietly shipped anyway.

• Your circadian phase, as a clock time. Published wearable phase estimators carry roughly ±2 hours of error and depend on light exposure, which no connected device measures. “Your melatonin onset is 21:47” is exactly the kind of claim that turns an honest layer into a wrapper with extra confidence. Velisa shows your sleep-timing midpoint instead, disclosed as not the same thing.
• A simulated Process S. The physiology is real, but its standard time constants are calibrated against EEG slow-wave activity no wearable provides, and the commonly quoted values could not be traced to a primary source. A transparent decaying sleep-debt heuristic ships instead, labelled a heuristic everywhere it appears.
• The coupled acute:chronic workload ratio most training apps ship, where the last 7 days sit inside the 28-day average. Mathematical coupling manufactures an apparent injury association even in randomly generated data, because the numerator is part of the denominator (Impellizzeri 2020). Velisa computes the uncoupled version — which is why its number will not match an app that couples.
• A body-battery figure. Beat-to-beat data across a whole day is not available from any source Velisa reads, and the minutes-to-points mapping is unpublished everywhere. It would be a reimplementation of an unpublished algorithm on worse data than the original.
• A precise South Asian body-fat threshold. The literature states the direction the correction should move and publishes no magnitude. Waist cut-offs are published and are implemented — 90 cm male, 80 cm female, against 102 and 88 for the global set — but the body-fat offset returns nil in code, and the gap is disclosed in the output rather than filled with an invented coefficient.

SourceVelisa/Scoring/References.swift, ReferencePopulation.unverifiedBodyFatOffsetPoints — a documented nil with a comment saying what would have to be published for it to become a number.