The research layer

The workings

How I know what I know, for the curious and the sceptical. Two questions run this whole layer: can you trust any of it at all, and how do I know I didn't fool myself? Every room below answers a piece of one of them.

The rule this layer enforces: the surface of this site has the liberty to simplify; it never has the liberty to overstate. Everything nearer the surface must trace back to what's here.

start here · The skeptic's tour Can you trust any of this? Take the tour → Six rooms, about fifteen minutes, the whole trust argument, walked in order: the spine, the definition everything keys off, the discipline, the honest ceiling, the limits, and the graveyard. 6 stops · ~15 min · leave whenever you like

In a hurry? The three load-bearing rooms: Three things that have to be true to make this research possible · A discriminator, not a predictor · What didn't survive , or browse every room below.

The foundation

Can you trust it at all?, what we measure, and whether it connects to anything

What we measure

Case definition & measurement

Before any finding: what the watch records, what counts as the thing we're trying to predict, and the four chapters the data falls into.

What counts as a crash → The definition everything keys off, a run of bad days, score-based, locked before any test ran. ~2 min · Outcome definition, a pre-specified, score-based case definition, locked before any analysis. What a 3 actually means → The felt-state score in lived terms, what each level is in the body, in the participant's own words, and why the ceiling is a 6. ~5 min · Construct anchoring, participant-authored lived-functioning definitions per score level; the phenomenological referent the crash case-definition rests on. Four years, four chapters → The record isn't one even stretch; it falls into chapters set by the illness itself, with a lived-experience recovery story inside the longest one. ~3 min · Temporal stratification, data-given strata (hard boundaries) + a lived-experience recovery-phase axis (M1-warranted, not data-driven boundary detection). The recovery, in six phases → The six stages I lived through, and a fair test of whether the boundaries I drew from memory show up in the watch data, without ever being tuned to it. ~5 min · Convergent validity of lived-experience boundaries, per-boundary distribution-shift (KS, Cohen's d) + data-driven change-point comparison across 7 channels (Q4.3). From watch to number → How a wrist sensor becomes a row of numbers the analysis can use, and the choices in between. ~2 min · Measurement pipeline, raw sensor samples to derived daily features; within-person (baseline-relative) scaling. What 'stress' really measures → Garmin calls it 'stress', but the number is a read of how activated the nervous system was, not a measure of mental stress, and blind to whether the cause was worry, a virus, or a hard walk. ~2 min · Construct validity of 'stress', an opaque HRV-derived autonomic-arousal index, not a validated psychological-stress measure; source-indeterminate (exertion, illness, pharmacology all load it). The data dictionary → What every variable behind the findings means, in plain language, the technical name and details one fold down. ~15 min · Variables & provenance, each channel, its derivation, units, and missingness. The guide, statement by statement → Every statement in Wiggers' guide, distilled: what it says, whether it is about crashes (PEM) or circulation (POTS), whether this watch can measure it and how, and a link to where I assessed each. ~6 min · Source-coverage ledger, the full Wiggers-statement register with mechanism labels (PEM/POTS/both) and per-statement measurability, cross-linked to the tested findings.

The discipline

How do you know you didn't fool yourself?, the honesty apparatus

How a claim is made

Pre-registration & study design

The discipline that turns a hopeful pattern into a finding you can trust, or refuses to.

How a question gets locked → Pre-registered, audited in a fresh session, and frozen before it runs. The discipline against fooling yourself. ~2 min · Pre-registration, hypotheses and analysis plan locked before data, fresh-session audited; guards analytic flexibility. Anatomy of one finding → One real signal walked end-to-end (claim, pre-registration, the null, the verdict, the stress-test) so the whole method is concrete. ~3 min · A worked example end-to-end, operand, block-permutation null, discrimination statistic, sensitivity arms.

How far a finding reaches

Discrimination, multiplicity & robustness

A real signal still has a ceiling. What a result does, and does not, entitle anyone to claim.

A discriminator, not a predictor → Why even the sharpest signal here is a weather report, not an alarm, base rate, precision, and the honest ceiling. ~2 min · Diagnostic ceiling, base rate, positive predictive value; retrospective discrimination vs prospective prediction. Stress, and how I felt → Does the watch's 'stress' have anything to do with how I felt? Why the autonomic science says it should, what shape to expect, why Wiggers makes me look so closely, and what a first descriptive look finds. ~9 min · Functional form & motivation, autonomic-arousal construct, PAIS dysautonomia priors, and the literature-predicted non-linear dose-response; near-zero linear ρ conceals a significant non-monotone curvature, felt-state peaking mid-stress (HA-C3, spline F=28.27); episode-matched z recovers the signal daily-linear measures cancel (Q4.9). Seven signals, or one? → Several signals looked like independent witnesses agreeing, until we checked. Most were the same signal, measured several ways. ~3 min · Multiplicity & collinearity, cross-channel correlation, effective number of independent tests, corrected significance. The COVID check → The one pre-registered test of whether the watch tracks a real, independently-known event: did its autonomic signal move during a 14-day COVID infection? A qualified yes, carried by the stress and body-battery channels, with heart rate masked by peak fitness. ~3 min · External-event validation, a pre-registered, no-peek peri-event test of the cross-channel autonomic factor against an independently-known infection window (block-bootstrap null; git-provable locked design; partial 2/3 coherence, RHR fitness-masked).

What else could explain it

Confounding & causal inference

The rival explanations for any pattern over time, and where each one stands.

What else could explain this? → The driver ledger (recovery, pacing, medication, season, chance) and where each candidate stands in the investigation. ~9 min · Confounding, time-varying covariates on an investigation-status ledger; the limits of causal inference at n-of-1. The citalopram question → One medication, one changing dose, three channels it demonstrably moves, everything we know and tested about the confounder sitting inside the main signals, in one place. ~7 min · Pharmacological confound, dose–response (β per mg) across three channels, with cross-window replication.

The limits, and the graveyard

Limitations & falsification

What this study can never tell you, what the instrument can't see, and the findings we tested and threw away.

The seven limits → The honest floor: the seven systemic limits a four-year, one-person study can never escape. Every finding inherits them. ~2 min · Limitations, single-subject (n-of-1 / single-case design), external validity, presence-only data, analyst-as-subject. What this watch can't see → The blind spots of a consumer device, no HRV, proxy signals, coverage gaps, and notes that record presence but never absence. ~2 min · Measurement validity & coverage, unmeasured constructs (no HRV), proxy signals, coverage windows, presence-only logging. What didn't survive → The findings that looked good and didn't hold up, kept on display. The clearest test of an honest site is what it throws away. ~2 min · Falsification, refuted hypotheses, baseline artefacts, and multiplicity deflation, kept on display. The danger window that wasn't there → A pre-registered test of the most useful thing a watch could catch (that pushing too soon after a crash tips you into another) came back unable to find it. Why an honest 'cannot resolve' is itself a result, and why nothing predictive ships from it. ~4 min · Pre-registered inferential null, post-crash danger-window push-relapse test; peak cardiac-strain-vs-matched-baseline relapse contrast, event-level block-permutation CI spans the null at n=24 event-units; 'cannot resolve' as the pre-committed reading; strengthens the no-visible-trigger finding. The closed-hypothesis ledger → The un-spun floor: the five discipline gates every result passed, and the seven pre-registered crash tests with their single-pool verdicts, one supported, six not. Plus links to the full record. ~2 min · Research-layer package, the pre-registration/review/privacy/single-pool gates, the seven-signal single-pool ledger (1/7 SUPPORTED) with usability metrics, and the un-abridged registry.

The road

How it actually went, the messy real path, kept honest

How the investigation went

Research narrative

The whole arc as a story, the questions, the rethinks, and the turns in method.

How the investigation went → The storyline of how the hypotheses evolved, the forks, the rescues, the dead ends, the reckoning. ~11 min · Research narrative, the hypothesis genealogy and the methodological turning points. The open-work register → Everything I raised past her guide and where each stands: the findings that graduated to their own pages, the shapes I can only describe, and the open questions still on the list, sorted by how settled each is. Showing the work in progress is the point. ~4 min · Beyond-the-guide open-work register, the personal-hypothesis (P-register) surfacing by kind × stage (descriptive/tested × idea/scoped/done/parked/inconclusive), work-in-progress inclusive; the beyond-side twin of the guide-coverage ledger.

The backbone

This layer is the source of truth for the whole site. Every claim nearer the surface (a finding I report, a line in the field guide) must trace back to, and stay consistent with, what's here. The surface has the liberty to simplify; it never has the liberty to overstate. A reader who descends this far should find each claim confirmed and given context, never contradicted.

What never ships, here or there: the raw personal data. The daily felt-state scores, the written notes, and the dated calendar entries stay private. Only aggregated, baseline-relative values are ever published, the privacy rule inherited from the research itself. The full pipeline, methodology, pre-registrations, and audit reports are already public at the repository; this layer is the readable, on-site version. Also live: reading & sources; and back up top, the story, the field guide, and beyond the guide.