Signals
What data the system reads — and what it cannot tell us.
Purpose
This document defines the exact data inputs available to the system. Every signal listed here comes from consumer wearables. No signal in this list constitutes a medical measurement.
A developer reading this page should know:
- What can be read from the hardware
- What can be derived with reasonable confidence
- What must never be inferred from this data
Supported Signals
| Signal | Source | Characteristics | Update Frequency |
|---|---|---|---|
| Sleep Stages | Wearable (accelerometer + PPG) | Estimated, not clinical polysomnography. Light/Deep/REM classification. | Nightly |
| HRV (Heart Rate Variability) | Wearable (PPG) | Trend-based, noisy, influenced by hydration, alcohol, position. rMSSD or equivalent. | Nightly / Morning |
| Resting Heart Rate | Wearable (PPG) | Delayed indicator. More stable than HRV. Lower is generally better for the individual. | Nightly |
| Sleep Consistency | Derived (bed/wake times) | Pattern regularity across nights. Circadian alignment proxy. | Rolling window |
Signal Characteristics
Trend-based, not absolute
All signals are meaningful relative to the individual's baseline, not as standalone values. An HRV of 35ms means nothing without context. An HRV that dropped 30% from a person's 90-day average is meaningful.
Noisy
Wearable sensors produce estimates, not clinical measurements. Single-night anomalies are expected and must not trigger alerts alone.
Delayed
Most recovery signals reflect what happened 12–48 hours ago, not the current moment. The system must account for this latency when interpreting data.
Influenced by confounders
Alcohol, caffeine, travel, illness, menstrual cycle, and dozens of other factors affect every signal. The system cannot isolate causality.
What Signals Cannot Tell Us
This is as important as what they can tell us.
| Impossible Inference | Why |
|---|---|
| Diagnose a condition | Consumer wearables are not medical devices. Signal accuracy is insufficient for diagnosis. |
| Determine causality | A drop in HRV does not tell you why it dropped. Correlation ≠ causation. |
| Prescribe treatment | No signal combination justifies medical intervention. |
| Replace a doctor | Even perfect data cannot replace clinical judgment, history, and examination. |
| Predict disease | Trends may correlate with health outcomes in research, but individual prediction is not validated. |
Population Averages Are Not Decision Inputs
The system explicitly rejects population averages as primary decision criteria.
Why:
- "Normal HRV" ranges vary by age, sex, fitness, genetics, and measurement method
- Telling someone their HRV is "below average" creates anxiety without actionable insight
- The only meaningful comparison is you vs. your own baseline
Rule: Population data may be used for initial context during onboarding but must never drive agent recommendations or deviation detection.
Developer Guidance
DO
- Compare signals to the user's personal baseline
- Require multi-signal confirmation before triggering agents
- Treat single-night anomalies as noise
- Display trends, not snapshots
DON'T
- Show "normal ranges" based on population data
- Use language like "your HRV is low" (low compared to what?)
- Infer mood, stress level, or health status from a single metric
- Present wearable data as medically accurate
Bottom line: These signals are behavioral feedback tools, not diagnostic instruments. Build accordingly.