What this watch can't see.

• No heart-rate variability

  The Forerunner 245 records no HRV at all: the textbook autonomic early-warning signal, and the one the pacing literature leans on hardest. It is simply empty for every date across four years. We worked around it with an overnight-stress proxy; the HRV question itself stays formally untestable on this watch, forever.

• 'Stress' is a proxy, not a measurement

  Garmin's stress score is an opaque, HRV-derived index: useful, but not a validated clinical measure of anything. We treat every stress-based finding as resting on a stand-in, and say so rather than dressing it up as a direct read of the nervous system.

• Some signals only exist recently

  Per-minute overnight Body Battery, for example, is only populated from about mid-2024. This cuts two ways. A signal that exists only lately can manufacture a 'recent-era' pattern that is really about when the data starts, not the body, so coverage windows have to be checked before any difference over time is read. And it runs the other way too: because the early, crash-heavy years lack this fine-grained data, a whole class of sharper tests (per-minute recovery, or the detailed sleep breakdown a proper sleep-precursor test would need) can't be run on the years where most of the crashes actually were. Those questions stay unanswerable here, not merely unrun. See the measurement-regime driver.

• Notes record presence, not absence

  A symptom written down on a given day is real evidence it happened. A symptom not written down is not evidence it didn't. The note record is presence-only, so a blank can never be read as proof that nothing occurred.

• The felt-state is a feeling

  The daily 1–10 score is an honest self-report, not a lab value, and it is the ground truth the entire site rests on. It is unavoidably subjective, and no amount of sensor data changes that.

• No labs, no second instrument

  No bloodwork, no validated sleep-stage scoring, no second device to cross-check the first. One wrist, one watch, four years. Anything that needs an external measurement to confirm it stays out of reach by construction.

None of this sinks the project; it scopes it. A signal read honestly through its blind spots is worth more than one that pretends the instrument saw everything.

The full sensor audit, in the research repo: garmin_indicators_audit.md ↗. The systemic version of these is the seven limits.