Ask the Twin: Auditing Counterfactuals of Wearable World Models

A digital twin aims to simulate an individual's physiology under hypothetical actions. Yet accurate forecasts of observed trajectories do not establish accurate counterfactual forecasts. We introduce a framework for auditing a wearable world model's counterfactuals using two design-based estimates from the same observational data: within-person and matched cross-person difference-in-differences. We use published randomized trials to anchor the expected population effect and evaluate the model's effect direction, magnitude, and variation across subgroups. For GLP-1 initiation, a wearable world model trained on 6.3M users recovers 35% of the observed resting-heart-rate effect, averaged over horizons of 1 to 30 days (observed day-30 effect 1.81~bpm). Curated post-training brings it to 87% and recovers the BMI gradient; on age it matches the observed effect in three of five strata.