The path the investigation actually took.
Findings are the clean version. This is the messy one: the order in which the understanding actually developed, dead ends and rescues included. It is the strongest evidence for honesty we have: we didn't get it right the first time, and you can see exactly where.
How the investigation actually went
The tests didn't run in a straight line, and we didn't get them right the first time. This is the real path: which questions we started with, where the body forced a rethink, where the watch's hardware blocked us and we improvised, and how a confident-looking finding got cut back down to size. And late on, a bigger correction still: we had been asking only half of the guide's question, so we went back for the other half. The messiness is the honesty, and the path isn't finished: new analysis is still landing, and this storyline is kept up to date as it does.
A storyline, not a calendar: what's sequenced below is the order in which the understanding developed, not the order (or the dates) the tests were actually run in. And it's curated for progress, not completeness: each card marks a genuine step in the thinking (a question we started from, a turn in method, an operationalisation or hypothesis that followed) not every test we ran.
Provisional first cut. The findings are placeholders that update as the research closes, and the ~40 internal hypotheses are simplified to the sixteen that carry the story.
How to read this board
- The crash hunt: Can the watch see a crash coming?
- The rivals: What else moved the numbers?
- The reckoning: How sure are we, really?
- The two axes: Was it even a crash question?
- refined: measured the same thing a sharper way
- revised: hit a wall, rewrote the rules, tried again
- rescued: the watch couldn't see it, so we found a proxy
- followed up: a result raised a new question
- sibling: the same pattern read on a different channel
- complicates: a confound: this may have changed the very signal we were reading
- corrected for: the confound is established and now modelled out of the signal's reading
- recast: the same finding, read on the axis it was actually about
- Not found Looked for it; it isn't there in my body.
- Showed up The pattern appears in my body, across the whole four-year record.
- Inconclusive The test ran but landed underpowered, too few crashes or too little signal to call. A real state, not a gap.
- Couldn't test directly The watch can't measure it; tested via a proxy, if at all.
- Act 1
Start with the famous rules
I began where anyone would: the best-known pacing signals, tested exactly as written. The most famous of them, the morning heart-rate rule, came up empty on my body. That first surprise set the tone, nothing taken on reputation.
- H01 Not found
Resting heart rate before a crash
It didn't show up. If anything, in the recent years RHR ran lower before a crash, not higher.
The story · how we read it
The Workwell Foundation's 'resting heart rate +15' rule, the most established early-warning signal in the pacing world.
We started exactly where the literature points. The most famous rule of all simply didn't hold for my body, which is what made the subtler signals worth chasing.
- H02 Showed up
Sustained stress before a crash
Promising in the early years, absent later. And a daily average turned out to be the wrong lens.
The story · how we read it
Was the body under sustained stress in the days before a crash, measured as Garmin's daily-average stress?
The first hint that something real lived in the stress channel. But averaging a whole day buries a single sharp moment, so this became the seed for a much sharper test later.
- HA06 Not found
A morning heart-rate swing
Refuted throughout. In absolute beats, the thresholds were mis-calibrated to my body's small range.
The story · how we read it
A gentler version of the heart-rate rule: any unusual morning swing, up or down, against a personal baseline, Wiggers' bidirectional framing.
Re-opening the heart-rate question with a softer, two-directional shape. It failed as posed, but pointed straight at a measurement fix.
-
- Act 2
The body kept changing
Partway in, the crash counts fell sharply, from about ten a year to about two. Real recovery, and worth being careful with: the body late in the record is not the body early on, so any change over time is part real, part smaller sample. A thing to note and hold lightly, not a spine to hang the rest on.
- turning point
The body kept changing
Real recovery, which the patient confirms, not a scoring drift; the body in the later years simply crashes far less than early on. It also thinned the recent data, fewer crashes to test against.
The story · how we read it
Partway in, the crash counts fell sharply: from about ten a year to about two.
The body late in the record is not the body early on, so an over-time difference in any signal is part real change and part smaller sample. Worth being careful with: a reason to read the whole record as one pool and to hold over-time stories lightly: a number, not a narrative, and not a spine the rest hangs on.
- K01 · K02 Inconclusive
Did the kind of crash change?
↰ followed up from The body kept changing
Both point the same way: no recent crash hit the very bottom of the scale where the early ones did, and the long-crash tail collapsed (episodes of five days or more fell from 36% to 7%). But with so few episodes, neither move quite clears the strict bar, inconclusive.
The story · how we read it
As the body recovered, did the crashes themselves get milder, both less deep, and shorter?
One step in the thinking, tested two ways (depth and duration) the question the cliff had just handed us: did the crash itself change? Both reads agree the recent crash is a milder animal than the early one; neither has the sample size to call it. (Two hypotheses, K01 and K02, folded into one step here.)
-
- Act 3
Sharper instruments
So the questions were rebuilt to fit my body: personal baselines instead of fixed thresholds, short spikes instead of daily averages, and a stand-in for the HRV this watch cannot record.
- H02b Showed up
Stress spikes, not stress averages
The early signal sharpened into the strongest in the whole project. Gone in the later years.
The story · how we read it
Not the daily average, the count of short, sharp stress spikes in a day.
The patient's own words reshaped the measurement: 'an intense moment in an otherwise calm day can trigger a crash.' A daily average dilutes a five-minute spike across the other 1,435 minutes. Counting the spikes instead is the single clearest methodological turn of the investigation.
Even this, the strongest signal we have, is one we can only read after the fact.
- HA06b Showed up
The heart-rate test, re-scaled
An early signal appeared once the thresholds were scaled to me.
The story · how we read it
The morning heart-rate test again, but in standard deviations from my own baseline, instead of a fixed number of beats.
A lesson banked as a standing rule: a threshold borrowed from a population has to be re-scaled to the individual before you test it. The 'relative, not absolute' principle was born right here, out of HA06's failure.
- HA07 Couldn't test directly
Heart-rate variability
Untestable on this watch. The Forerunner 245 simply doesn't record HRV; the data is empty for every date across four years.
The story · how we read it
HRV, the textbook autonomic early-warning signal, and one Wiggers leans on directly.
A hard wall. The most theoretically promising channel of all was invisible to the hardware. Rather than quietly drop it, we asked what else might carry the same information.
- HA07c Showed up
A stand-in for HRV
The proxy carried the early signal cleanly.
The story · how we read it
Garmin's overnight 'stress' score is itself HRV-derived, so we used sleep stress as a proxy for the HRV we couldn't read.
The rescue. A blocked hypothesis became a substitute one, pre-registered before we looked at the data, so the discipline held. The HRV question itself stays formally untested forever on this dataset; its proxy does not.
-
- Act 4
The strange thread
The oddest part. Some nights read as perfectly recovered, and the crash came anyway. The question turned from how loud the signals were to how still they went, and the one signal that held across the whole record lives here.
- HA10 Showed up
How recharged the morning looked
The first signal to show up in the later years, and it ran backwards: the body looked more recharged before a crash, not less.
The story · how we read it
How recharged the body looked by morning, from Garmin's Body Battery.
The turn in the story. After thirteen straight tests found nothing in the recent era, here was one, and it pointed the opposite way to the textbook. A paradoxical 'swing': the night looked great, and then the crash came anyway. (The reckoning, later, shows this is the lay-readable face of the same overnight state HA07d catches as a movement-collapse, one signal, two readings.)
- HA07d Showed up
When the body goes still
↰ refined from HA07c ↰ sibling from HA10
Before recent crashes that variability collapsed: the body went unusually still. The one finding that held across the whole record.
The story · how we read it
Not the level of overnight stress, but its variability, how much the autonomic state moved from night to night.
The emotional centre of the whole investigation. The recent-era crash doesn't look like over-arousal; it looks like a freeze: an autonomic stillness that reads like wonderful recovery while the symptoms keep coming. Arrived at purely from the data, it lands on Wiggers' own parasympathetic-'swing' pattern. Convergent discovery, not an audit.
Provisional: it cleared our pre-registered bar and held up across the whole record, though with one body the uncertainty stays wide.
-
- Act 5
The rivals
Before believing any of it, the other explanations had their turn: the treatments running through the same four years, and the marks they may have left on the very channels being read.
- intervention
Citalopram and the stress channel
Confirmed to move the watch's autonomic channels in a dose-graded way, and counterintuitively: more citalopram nudged measured stress UP and the body-battery floor DOWN. (The likely-but-unproven why: SSRIs can blunt heart-rate variability, which the watch reads as higher stress, a literature review found that mechanism mixed, not settled.) A medication fingerprint in the data, not the body's state. That makes it a real confounder of any raw change over time, now modelled out of the signals it touches, where subtracting it barely moved them and nothing grew stronger.
The story · how we read it
Citalopram (an SSRI) was one of the treatments running across these four years. We asked whether it left a mark on the data itself.
A genuine complication, not a footnote, and the investigation's clearest case of the watch measuring the drug, not the person. Stress and body battery are exactly the channels the precursors are read from, and the dose changed across the four years (none early, 30mg through the recent era, tapering now), so any 'change over time' in them is part body, part drug. So we subtracted the medication before reading recovery, and it barely moved the change-based signals. Everything we know and have tested on citalopram is pulled together in the workings, under 'The citalopram question'.
The dose-response itself is confirmed (three channels, three converging tests), and it's now been used to correct the signals it touches, a small correction that strengthened nothing. With one person it stays 'consistent with a dose-graded response, after the recovery slope is absorbed', not the drug proven to act alone.
- intervention
CPAP and the overnight signals
Under analysis. Listed here so the thread is visible, not hidden: the overnight channels (sleep stress, body battery) are exactly the ones a breathing intervention could move.
The story · how we read it
CPAP (for sleep-disordered breathing) was another intervention across the window. Did it change the overnight numbers the autonomic tests lean on?
Part of the same honest question as citalopram: before reading an overnight signal as the body, rule out the machine helping the body breathe.
In progress, no verdict yet.
-
- Act 6
The reckoning
Two hard questions the research put to itself: were the surviving signals really separate, or one signal measured several ways; and would even the best of them actually warn anyone. The claims got smaller, and firmer.
- the reckoning
Seven signals, or one?
Cross-channel correlation collapses them into three or four clusters; corrected honestly for that, only one (the stress spike) cleared conventional significance.
The story · how we read it
We had a handful of 'supported' early-era channels. But were they really separate?
The deflation we ran on ourselves. 'Many converging channels' was overstated: several were near-identical signals wearing different clothes. The honest headline is smaller than the exciting one, so we changed the headline.
- the reckoning
Would it actually warn you?
Under one time in twenty. A warning would be a false alarm nineteen times out of twenty. These signals read the body; they do not predict it.
The story · how we read it
Take even the best signal. If it fired tomorrow, how often would a crash really follow?
The line that governs the entire site. A signal can clearly separate crash-weeks from normal-weeks in hindsight and still be useless as an alarm. That is the honest ceiling: a weather report on the body, not an alarm.
-
- Act 7
The other half of the guide
For most of this, we ran everything through one lens, prediction: can the watch see a crash coming? We leaned on that hard, and, honestly, we overdid it. Only once most of that work was done did we take the time to give the other half of Wiggers' guide its due, the POTS side, circulation and standing and blood volume, which was never about crashes in the first place. Read that way, some of the patterns we had put through the crash test simply sit on a different axis, present and real, and not about the next crash at all.
- the reckoning
The other half of the guide
The guide is really two guides: a crash axis (PEM), where a prediction test belongs, and an orthostatic axis (POTS), which is descriptive and about the body in the moment. In my data the two mostly fall on different days.
The story · how we read it
Wiggers has ME/CFS and POTS. Much of her guide is about the POTS side, circulation, standing, and blood volume, which is not about crashes at all.
For most of the investigation we ran everything through one lens, prediction: can the watch see a crash coming? We leaned on that hard, and, honestly, we overdid it. Only once most of that work was done did we go back and give the other half of the guide its due. It turns out to be about half of what she writes, and it was never a crash question, so the rest of the site now follows the split.
The two-axis split and its numbers are early, descriptive, and still under review; the framing holds, the exact figures are held lightly.
- HA11 Showed up
Within-day 'U-dips'
↰ recast from The other half of the guide
Present in my data, and time-varying, more in the early years than the recent ones, though those recent years are also when the medication began, and I never took up the salt-and-fluid management she describes. The circulation pattern Wiggers names, not a crash precursor.
The story · how we read it
A sharp drop in stress followed by a rebound higher than before, an orthostatic pattern Wiggers describes, counted within each day.
The one orthostatic sign this watch can see. Wiggers ties it to blood volume and treats it with electrolytes; in my data it shows up and comes and goes over time, which is how a circulation pattern behaves, not a warning before a crash.
-