The COVID check.

1. 1
   The problem

   One link needed evidence from outside the watch

   The whole site rests on a chain, and its first link (does the watch see anything real?) is the one leaned on most lightly. Everything else on this page could be true and the watch could still be reading its own noise. To answer it, you need a moment where something real and independently-known happened to the body, and you can ask: did the watch move for it?

   The three things that have to be true →
2. 2
   The event

   A 14-day COVID infection: the one known autonomic event

   Across four years there is exactly one autonomic event whose timing is known from outside the watch entirely: a documented COVID infection, a 14-day window. An acute infection is a textbook autonomic stressor. So the test writes itself: did the overnight autonomic factor depart its ordinary baseline during that fortnight?

3. 3
   The lock

   The test was frozen before it ran, provably

   This is the part that makes it trustworthy. The whole design (the factor, the window, the null, the bar) was pre-registered, cold-reviewed in a fresh session, locked, and committed to version control before the test was executed. The git history proves the design predates the outcome; two separate fresh sessions guarded the no-peek contract. There was no room to tune the test to a result already seen.

   How a question gets locked →
4. 4
   The result

   It moved: external corroboration, qualified

   The overnight autonomic factor sat at the high-load pole during the infection window: the 99.2nd percentile of ordinary fortnights, a one-sided p-analogue of 0.0077, an effect around 0.86 of a standard deviation. The watch's factor moved, significantly, for an event it was never shown. That is the external corroboration the first link promised.

5. 5
   The honest nuance

   Two of three channels moved as predicted, and the third is explained

   It was not a clean three-for-three. Overnight stress rose and morning body battery fell, exactly as an infection should push them. But resting heart rate did not rise on the raw scale; it only rose after removing a downward pre-illness trend. That is explained, not a failure: right before COVID I was at my athletic resting-heart-rate floor, and that peak fitness masked the infection's heart-rate rise. So the result is carried by the stress and body-battery legs (which are the same autonomic factor read two ways) while the heart-rate leg was hidden by how fit I was at the time.

   Fitness history, in the driver ledger →
6. 6
   What it does and doesn’t earn

   A qualified pass: corroboration, not proof

   Held tightly: this is one body, one event, and the infection window abuts the onset of Long COVID, so the honest statement is that the factor departed baseline around the infection / LC-onset hinge, never that the infection caused it. It stays a strictly descriptive check of whether the factor tracks a real event; it makes no claim about the factor's overall validity. And the resting heart rate involved is normal in absolute terms; the signal is in the trend, not an alarming number. A qualified pass: the first link's lightest-leaned-on check, delivered.

   Seven signals, or one? →

A clean "it moved" softens nothing that was over-promised; it delivers the check the first link always said it was still waiting on. The infection shows up in the record, through the channels that share one autonomic voice, while the most independently-trustworthy channel was masked by how fit I used to be. Both halves are true, and both are on the page.

The pre-registration, the locked design, the test, and the result, in the research repo: peri-event-covid/result.md ↗