One number, or the whole night

Journal · 2026-09-18 · R. Klug · All notes

Your heart beats roughly 25,000 times while you sleep. A band on your wrist detects nearly all of them. In the morning, most apps hand you one number.

That number is not wrong. It is just almost everything thrown away — and the part thrown away is the part that tells you whether the number deserved your trust.

What a night looks like when you keep it

Here is one night from my own record, as ampulse stores it:

During training the same principle applies at a different speed: heart rate every second, every beat interval alongside it, and an explicit list of the moments the signal dropped out — rather than a smooth line drawn across the gap.

Why the quality of each beat is not a detail

This is the part that changed how I think about recovery metrics. HRV is not gently degraded by noise — it is destroyed by rare mistakes.

A study by Bourdillon and colleagues (2022) tested this directly. Adding fifty milliseconds of random jitter to every single interval in a recording changed RMSSD — the standard HRV metric — by zero percent. Then they inserted one realistic missed-beat artifact. RMSSD rose by 413 percent.

One bad beat out of thousands can quadruple your morning number. Which means a recovery figure without any statement of measurement quality is not a measurement — it is a number with an unknown chance of being nonsense. And optical sensors, all of them, produce artifacts: a wrist shifts, a band loosens, an arm goes under a pillow.

So the question is never whether your sensor makes mistakes. It is whether the app tells you when it did.

What this makes possible

A record kept at full resolution can be re-examined. When a value looks strange, you can open the window it came from and see whether the sensor was struggling — or whether that was genuinely your physiology. I have caught oddities in my own nights that I would never have found in a summary, simply because the underlying intervals were still there to inspect.

It also means the record outlives today's algorithm. Methods improve; a stored sequence of intervals can be re-analysed under better ones in five years. A score computed tonight is frozen at tonight's understanding.

And it means the record is portable. Export the intervals as FIT, CSV or a Kubios-compatible file, or let your own AI assistant read them through a connector that runs on your machine. The analysis is not something we rent you — it is something the file lets anyone do.

Where the industry stands

For context rather than comparison: WHOOP's developer documentation, as published in September 2026, states plainly that continuous heart rate data is not available through its API — the night comes back as summary values. Fitbit's own documentation would have made the same point, but the page returned a 404 while this note was being written — its developer API was retired in September 2026. The successor, Google Health, lists daily and interval heart-rate-variability types, and no interbeat interval anywhere in the catalogue. The better cloud products return five-minute samples, which is genuinely useful and a real step up.

Beat-level data exists in that world too — but reaching it means a commercial licence, a partner agreement, or a retention window measured in weeks. None of that is villainy; it is what happens when a measurement has to travel through a company before it reaches the person it came from.

Your physiology, at the resolution it was measured

ampulse is built on one decision: keep what the sensor produced, attach what is known about its quality, and store it where you can reach it. Everything the app shows is computed from files you can open yourself.

Check the math yourself. That sentence only means something if the math is still in the file.