The recovery, in six phases.
Descriptive · in progress · the phases are set from lived experience; what the data shows across them is still being read
I didn't live these four years as one long stretch. I lived them in stages, and the boundaries between those stages come from memory, from real events in my own recovery, not from the data. Which sets up a fair test: do the lines I drew from life show up in the watch's numbers anyway?
The six phases
Six phases, two inherited from the illness's hard chapters and four from the recovery itself. Each boundary is anchored to a remembered event (an infection, the start of ergotherapy, the start of medication) so none of them could be quietly moved to flatter a result.
- 1
Before, a healthy baseline
Aug 2021 – Mar 2022A healthy body, wearing the watch already. The closest thing the record has to 'normal'.
- 2
The infection
Mar – Apr 2022Two weeks of acute COVID, the hinge between the two bodies the record holds.
- 3
Long COVID, before pacing
Apr – Sep 2022The months when the acute symptoms hardened into chronic patterns. No pacing yet, no treatment, just learning the illness was going to stay.
- 4a
Learning to pace
Sep – Nov 2022Ergotherapy began. The first eight weeks of onboarding, learning the practice and starting to build the habit.
- 4b
Pacing as a habit
Nov 2022 – Apr 2024The long stretch where pacing was established practice, though how well it worked was still pushed around by workload, season, and the illness itself.
- 5
On medication
Apr 2024 – nowCitalopram enters and measurably changes the autonomic numbers; the dose builds, holds, then tapers back down.
Does the watch agree?
Each boundary above was placed from memory, before anyone looked at the data. Then we checked: at each transition, how many of seven watch signals show a genuine shift in their distribution? The honest test of a lived boundary is whether the instrument, asked independently, sees the same seam.
Every transition I lived through shows a shift on at least four of seven channels, and not one boundary was placed by looking at the data.
- Into the infection Mar 20224/6
- Into Long COVID Apr 20224/6
- Into learning to pace Sep 20224/7
- Into pacing as a habit Nov 20224/7
- Onto medication Apr 20245/7
Run the other way, it holds too. Let an algorithm hunt the seven channels for change-points with no knowledge of my history, and it finds 108 of them. Eighteen land squarely on a boundary I'd drawn from memory, independent corroboration. The other ninety don't, and that's expected: each channel has its own drifts and quirks, moving for all sorts of reasons that have nothing to do with the recovery.
The shape of each phase
The meters above count the seams; this looks at the phases themselves. Pick a signal, and see its level across all six phases at once, and, at each seam between them, whether the watch registered a real shift.
Garmin's overnight stress, the HRV stand-in.
Overnight stress: a distribution shift at 5 of 5 lived boundaries. Every seam this channel could be tested at shows a shift.
box = middle half (IQR) whiskers = full range shift at this seam no shift dose-mixed (dashed)
| Phase | Before | Infection | Before pacing | Learning | Habit | Medication |
|---|---|---|---|---|---|---|
| median | 15.9 | 20.3 | 19.2 | 20.0 | 19.5 | 18.9 |
Box = the middle half of the days (p25–p75); whiskers = the full observed range (not statistical fences); line = median. n is small in the infection and learning phases, so those boxes are illustrative.The dashed phase-5 box mixes dose states (0–30 mg and taper), so read it as spread, not a settled level.
Read the levels with one caution. Most of these are shown as they were, not adjusted for everything else that changed across four years. The physiological lines ( resting heart rate most of all, and the stress and body-battery channels built from the same heartbeat timing) sit on a slow drift: over these years I gained about 15 kg, lost fitness, and aged. So a level gap between an early phase and a late one is part illness-phase and part that drift, and can't be read as the recovery alone. The one line that's clean is the felt-state score; that's my own report, immune to how fit or heavy I was. For the most exposed line, resting heart rate, I did the correction: pick that signal and a second, paler box appears beside each phase with the modelled drift (weight, fitness, age, medication) removed. The recorded level climbs into the later phases; the drift-removed level sits flat at about 52–53 bpm throughout, so that rise is the drift, not the illness. It's a bracketed estimate, not a clean split (the drivers are named and sized in the driver ledger), but no honest version of it recovers into a rising illness signal. The stress and body-battery lines aren't corrected the same way: their phase-5 step is mostly the medication, a separate story told in the citalopram question.
What this does, and doesn't, claim
- The boundaries were never tuned to the data. They come from lived experience; the data is a check, not the source. Drawing a line on the data and then testing it against the same data would be circular, so we don't.
- A quiet boundary is not a wrong one. Where the watch shows little at a seam, the lived transition is still real, the instrument just may not register that kind of change on these particular channels.
- This is descriptive, and in progress. It characterises where the lived map and the data agree; it makes no causal claim, and recommends changing no boundary.
This is the cleanest kind of agreement: a map drawn from lived experience and a map drawn from sensor data, made independently, lining up at most of the same seams. These lines were drawn by the illness and by memory (not chosen to fit a result) and the data simply confirms it could see them too.
The chapters these phases sit inside: four years, four chapters. What changed across the phases: the driver ledger.
The full backstop, in the research repo: Q4.3 era_boundaries · findings.md ↗ · The recovery-phase axis (methodology) ↗