UCLA

Deep sequence prediction models always achieve high prediction accuracy by using Continuous Glucose Monitor (CGM) data to conduct forecasting on glucose. However, the top accurate models always make serious ranking mistakes on causal relationships. When a counterfactual insulin or carbohydrate dose is applied, they always produce false rankings as the result. Thus, these models cannot be used as support for insulin dosing decisions. This paper studies how to acquire a prediction model that is both accurate and ensures causal validity. Thus, we built a multi-agent searching system by porting the AI Co-Scientist architecture of Gottweis et al. (2026) to glucose forecasting. Different from predicting glucose directly by using language models, these multi-agents propose, discuss, evolve, and evaluate stacking strategies on a fixed pool of pre-trained foundation models. Each strategy's "fitness" is its real held-out error, computed by fitting on a training split, selecting on a validation split, and reporting once on a test split, subject to a hard, benchmark-based causal gate. During the searching period, a veto-blend structure was discovered. In this weighted combination, a causal-safe pool of models, which includes a hybrid neural ordinary differential equation model and several tuned H2NCM variants, holds the majority, in order to ensure correct counterfactual ranking. Meanwhile, a high-accuracy black-box model only holds a small portion of the weight, which is used to improve prediction accuracy. This paper defines "causal-safe" as when the intervention ordering gap that a model observes is zero on a specified held-out benchmark grid. On the main Anderson cohort, the veto-blend structure achieved a 16.75 mg/dL root mean square error (RMSE) on a 30-minute prediction, with an 11.47 mg/dL overall RMSE, while maintaining zero observed causal error on the intervention benchmark. Compared to the current finest causal-safe model, whose results are 17.15 / 11.85, the veto-blend closed roughly 80% of the overall RMSE gap and 67% of the 30-minute gap to the causally unsafe black-box ceiling. The same mechanism replicated on the other two cohorts as well. On the Aleppo grid, the veto-blend is slightly better than the causally unsafe black-box model ceiling on both benchmarks. On the Brown cohort, its overall RMSE also reached the ceiling, while the RMSE on the 30-minute window still remains a small gap. On both cohorts, the model kept zero observed causal error on the intervention benchmark. Later, this paper will illustrate the reason why the remaining gap cannot be further closed. Recombining a fixed pool of strongly correlated causal-safe models saturates, and, among the levers we tested, additional architecture, recipe, or search effort did not measurably shift the frontier, whereas increasing the number of training subjects did. The major contribution of this paper is proposing a reproducible, grounded multi-agent method for building causal-safe forecasters, and providing a careful account of the price of causal validity in glucose prediction.