How much does recovery high min have to change before it stops being noise? The threshold from 730 nights

In 730 nights measured on a single body, recovery_high_min must shift by at least 79.7% before that change clears the noise floor at a 28-night window. That number seems extreme. It follows directly from the data. The coefficient of variation for this metric is 115.03%—meaning nightly fluctuations exceed the metric's own central value. What the dashboard displays as a clean number is, underneath, a variable with more noise than signal in any single reading.

What the device records—and what it omits

Recovery_high_min counts the minutes a body spends in a high-recovery state during the night. The number appears on the dashboard with the same visual precision as heart rate or temperature. The dashboard does not show the error interval around that number.

Over 730 nights on a single body, the median of this variable is 30 minutes. The coefficient of variation is 115.03%. At that level of dispersion, the variable fluctuates more than its own central value. One night can show 12 minutes. The next night can show 56. Both readings can be equally representative of the baseline state.

The autocorrelation coefficient—how much today's value predicts tomorrow's—is 0.273. That is close to zero. Last night's value does not explain tonight's value. The variable has almost no memory and a great deal of noise. Two traits the dashboard does not declare, but the dataset records precisely.

The gap in any single reading

Here is the problem that no dashboard surfaces: you see a number, but you do not see the confidence interval around that number.

If recovery_high_min drops from 56 to 30 minutes, the app may register it as a significant drop. But the threshold is not set by the app. It is set by the noise of the system itself.

The dataset measured that noise across 730 nights on the same body, with the same instrument. The output is called rcv: the residual coefficient of variation. It is the dispersion that remains after removing the long-term trend. For recovery_high_min, the rcv is 318.85%.

That is the gap. Not missing motivation. Not missing discipline. Missing a measured criterion for when a change is real and when it is still system noise.

The threshold table—from the source row

The dataset row for recovery_high_min contains four thresholds. Each was calculated over a different observation window:

Window	Minimum threshold to clear noise
28 nights	79.7%
56 nights	56.36%
84 nights	46.02%
168 nights	32.54%

The threshold is the minimum change an intervention must produce—given that observation window—for the effect to clear the system's noise floor. It is not a guideline. It is the result of 730 nights, 18483 measurements, and 31 recorded metrics on a single body.

There is an important asymmetry. At 28 nights, you need a 79.7% change. At 168 nights, you need a 32.54% change. The longer window does not demand a larger change. It demands a smaller one, because it has more nights to separate signal from noise. Time is not waiting. Time is precision.

Why a CV of 115.03% changes everything

When daily variability of a metric exceeds its own median, each individual reading carries a wide error margin. That does not make the metric useless. It makes the metric useless reading by reading.

The correct way to use a variable with a CV of 115.03% is with windows, not single points. Checking whether recovery_high_min was 30 or a few minutes higher last night carries no useful information. Checking whether the average of the last 28 nights shifted by 79.7% or more relative to the prior period—that has discriminating power.

A change below the threshold for the chosen window is not proof that a protocol fails. It is proof that the window is too short for this variable. Those are two different diagnoses, and confusing them leads to opposite conclusions.

The transfer

You are looking at the same gap that made the large recovery studies impossible to repeat. They measured group averages, not individual variability. We know high recovery matters. We do not know how much noise sits in the reading of that recovery for any specific body, on any specific instrument, over any specific period.

The 730-night dataset closes that gap for one body. It does not close it for yours. What it delivers is the model: how to measure the noise, how to calculate thresholds, what window each variable needs before its verdict means anything. The next step is applying that model to your own data.

The question that changes everything is not whether recovery_high_min moved. It is whether it moved enough to clear the threshold for the window you chose.

The limit

This data describes the noise in recovery_high_min as measured by a specific instrument, on a single body, over 730 nights. It does not say:

• Whether that level of variability holds for other bodies.
• Whether the device you use records recovery_high_min with the same algorithm or sensor.
• Whether reducing the noise floor of this variable is possible with any protocol.
• Whether a 79.7% threshold at 28 nights is high, low, or typical compared to other populations—that comparison does not exist in the published record.

The thresholds are not targets. They are the minimum needed to distinguish signal from noise. Clearing the threshold does not confirm that the change came from the variable you modified. It only confirms that the change exceeded the system's noise floor.

The most important limit: one body, one device, 730 nights. A second body may have a completely different CV. Data from a second subject measured with a different instrument cannot be combined with these results. The device is not comparable even when it uses the same name for the variable. This rule came from an internal error—when devices served as keys for subjects, one person's data overwrote another's without warning. Publishing the lesson is obligation, not confession.

The question this piece does not answer

What protocol produces a change of 56.36% or more in recovery_high_min over 56 nights? That is the next file. This piece delivers the criterion. The next one looks for whether anything meets it.

---

Mental Work — all protocols · related file: magnesium-for-sleep · open science — the dataset