EpiCoach is a system that treats a single athlete—here, a cross‑country mountain biker—as the subject of accountable N‑of‑1 science. The system records every claim, prediction, and outcome in an evidence‑gated registry, allowing each test to be audited or falsified.

The engine behind the predictions is a wrist‑only mechanistic digital twin. It integrates a lagged heart‑rate response, a fitness‑fatigue model with learned time constants, and a latent skill state inferred from repeated trail segments. For each training week, the agent prescribes a constrained experiment that maximizes information gain while staying within load and safety limits.

In a rolling‑origin replay benchmark using eighteen months of data, the twin outperformed standard forecasters for heart‑rate prediction but not for segment‑time prediction, and the agentic loop’s performance matched a static model.

Figure 2 shows these results graphically: the twin’s heart‑rate CRPS of 0.050 [0.042,0.065] surpasses the persistence baseline at 0.052, whereas for segment time the twin’s CRPS of 0.050 is essentially identical to the static twin’s 0.048 and both are notably higher than the 0.057 achieved by the best standard forecasters. A live prospective phase ran the weekly cycle and produced 20 predictions, of which 14 were scored. Forecasts for descent time calibrated after one correction, while climb predictions were initially pessimistic as the athlete improved. A safety officer vetoed the agent twice during the first three weeks. Each correction is logged in the registry, providing a template for auditable personalization in digital health.

However, after eighteen months of observational data, the fitness‑fatigue component’s time constants remain poorly identified. The model is ill‑conditioned, and time constants may be confounded with seasonal effects. Information‑directed design—planned deloads followed by standardized tests—could resolve these uncertainties, but such contrasts are not evident in the current observational schedule.