Please find my comments in the attachment. 

The manuscript presents an innovative ensemble strategy that combines a differentiable process-based model (δHBV) with a data-driven Long Short-Term Memory (LSTM) model, further diversified through the application of multiple meteorological forcing datasets. The approach is evaluated across a wide range of generalization scenarios (temporal extrapolation, PUB, and PUR) using the CAMELS dataset. Although the paper is well-written and the main ideas are clearly communicated, it would benefit from additional details in the methods and a deeper discussion of model complementarity and limitations.

Strength

1. The proposed ensemble framework is conceptually strong and offers a well-justified combination of complementary data and algorithmic modeling paradigms.
2. The study is well evaluated across well-defined training protocols and temporal-spatial splits, which improves confidence in its generalizability.
3. The use of multiple data sources for meteorological forcings addresses input uncertainty better than traditional single-source modeling.
4. The results, specifically the finding that δHBV improves spatial generalization, have clear implications for prediction in ungauged regions.

Weakness

1. Interpretability: While the δHBV model’s performance is shown to be beneficial in spatial generalization, the underlying reasons for this complementarity (e.g., structural constraints, parameter smoothness) are not deeply explored. A discussion of how each model contributes to ensemble diversity would strengthen the scientific value of the work.
2. Robustness and Sensitivity Analysis: The paper lacks an explicit assessment of how ensemble performance responds to errors or biases in the forcing datasets or uncertainty in model parameters. Including even a limited robustness analysis would improve confidence in the ensemble’s reliability. Additionally, the authors should consider running one or two experiments to understand whether changing the size of the lookback window (i.e., the number of historical timesteps) for the LSTMs impacted the overall performance of the ensemble.
3. Scalability and Practical Deployment: The manuscript does not address the computational or operational feasibility of deploying this ensemble framework in practice, especially over large domains or in real-time forecasting contexts. A short discussion (1-2 sentences) on this topic would add practical relevance.