The Science
The Signal Score is deterministic fusion, not a black box. We use Kalman filtering and CUSUM to detect longitudinal decline early — and we tell you exactly which parts are computed, estimated, or merely narrated.
Domain signals are normalized, weighted, and combined by a fixed formula. The same inputs always produce the same Signal Score — independently verifiable.
A Kalman filter smooths noisy day-to-day signal into a stable state estimate. CUSUM accumulates small deviations and raises an early-decline alarm the moment the trend breaks — weeks before a person would self-report feeling off.
At week 10, CUSUM crosses threshold — the estimated state has drifted enough to flag early decline. This is an estimate with a confidence band, not a diagnosis.
Every figure in the app carries its epistemic status. You always know how much to trust a number — and where the model's judgment begins and ends.
Deterministic
DeterministicComputed by a fixed formula. Reproducible to the bit. No model judgment.
Signal Score · sub-scores · R/A/G threshold
Estimated
EstimatedStatistical inference with a stated confidence band. Updates as data arrives.
Kalman state estimate · CUSUM change-point · VO₂ est.
Narrated
NarratedPlain-language explanation of the numbers above. The LLM describes, it never decides.
Monthly summary · trend narration
The LLM only narrates. It turns the deterministic and estimated numbers into plain language. It cannot change a score, move a traffic light, or raise an alarm — those are computed, not generated.