This paper describes recent updates to the WaterGAP global water resources model (v2.2e) and provides benchmarks against observations. 

As with previous versions, the model is impressive and the analysis is comprehensive. However, the improvements described in this paper, and their impacts on performance, appear fairly minor. There should be space in GMD for efforts that focus primarily on software and data updates rather than scientific ones, but revisions are needed.

1. The improvements include the treatment of reservoirs, updates to datasets for reservoirs, non-irrigation water use, and streamflow stations for calibration. Updates to model capabilities include PET representations, glaciers, and water temperature, though the latter has already been described in a previous publication. While all of these changes are justified, and no doubt an intensive effort, their impact on the modeled global water balance and distribution of NSE presented in the results does not seem to be a large change from previous versions of WaterGAP. Section 7.5.1 reports a nearly identical performance to the previous version. The scientific contribution of the new model capabilities should be more strongly justified.

2. A main focus of the updates is the reservoir model. From the previous paper (2021), the release policy is assumed to follow Hanasaki (2006) and Döll (2009), which distinguishes between irrigation and non-irrigation reservoirs. This is a simplified rule that can be applied globally, but is often inaccurate at the level of individual reservoirs. The current paper does not investigate changes to this assumption, but removes a previous limitation about the maximum storage capacity for flood prevention. The accuracy of reservoir storage shown in Supplemental Figures S1, S2 from version 2.2c leaves much room for improvement, and it is not clear that removing the storage capacity threshold will fix this. The updated results after the change are not shown.

3. More information should be included about the potential scale mismatch between reservoir outflow (a point) mapped to a larger grid cell. The same goes for the stream gage data used for calibration (Step 3 in Section 2.5.2). It is possible this information is included in previous papers, but it would help to discuss here the potential impacts of this scale mismatch.

4. The calibration process attempts to find an optimal value of gamma, the runoff coefficient, to align the modeled mean annual streamflow within either 1% or 10% of the observed. Failing this, additional correction factors are applied to the runoff. While I can appreciate the difficulty of calibrating a global model, this calibration setup would have several problems for a basin-scale study. The gamma parameter can compensate for any mass balance error without a physical relationship to the runoff curve shown in Figure 3 of the 2021 paper. The additional correction factors only worsen this problem, and many regions of the model rely on these (Fig 4 of the current paper). By calibrating to mean annual data, monthly dynamics could be lost, though the efficiency metrics reported in Figure 7 seem to be doing well at many stations.

This calibration approach may be standard for global models. But at the basin scale, we could expect to see more diagnostics applied to investigate whether the results are physically based, or to analyze how much of the calibration uncertainty comes from each component of the mass balance. The calibration is more of a bias correction that is not able to distinguish between the many degrees of freedom in the model.

The manuscript presents the most recent version of the WaterGAP model (v2.2e). Refinements, new algorithms, and updated data for calibration are presented. The manuscript is well-written, and the overall evaluation section is comprehensive and competently performed. My primary concern is the link between the extensive software modifications and data increment and the overarching scientific goal of assessing global and regional water resources. Even though the modifications, data additions, and model evaluation are comprehensively described, their scientific significance is not sufficiently justified in the context of water resources assessment. In this sense, I consider revisions (mostly related to rephrasing and further discussions) to be needed before publication. Please see below for my specific comments:

Comment 1: The authors cite three papers to introduce the WaterGAP model. However, definitions necessary to understand the implemented modifications are missing. As a reader, I spent extra time reviewing the referenced papers to understand the main changes. Examples of these definitions include "naturalized mode" (L45), "standard runs" (L46-L47), and "local lakes algorithm" (L46), to name some just in section 2.1. Introducing the model's main features might be better so readers don't need to jump between papers (see related comment 2). As for GMD guidelines, for model description papers, it should be possible for independent scientists to build a model that, while not necessarily numerically identical, will produce equivalent results. In the current manuscript stage, the last is very hard to achieve.

Comment 2: The model modifications are widely discussed in Section 2. However, as a reader, it is hard to visualize the model as a whole and identify the modified components. A scheme presenting the model's overall structure with the modified stages highlighted might be helpful.

Comment 3: Following an analysis of the simulated monthly time series of reservoir water storage to observations for 16 reservoirs in the United States, the authors decided to drop off the 85% maximum storage capacity assumption implemented in version 2.2d (Müller Schmied et al. 2021) (due to model underestimation in 11 of the analyzed cases). This implies that the decision to remove the original assumption is based on a local-based analysis where ~30% of the model results did not show underestimation issues. Thus, I am failing to see the reason for extrapolating these local results to the global implementation of the model. Furthermore, Figures S1 and S2 might indicate more pressing issues related to the model's accuracy (e.g., seasonality) that might not be improved by simply removing the assumption.

Comment 4: Section 5 presents the effects of the model modifications on multiple areas and the impact of differential forcing. Overall, the effect of the implemented changes seems to be minor. However, I found it challenging to visualize the minor differences because most of the baseline results (from v2.2d) are presented in Supplementary Material. I would recommend summarizing the results as differences of v2.2e from the original implementation rather than presenting each version independently. Furthermore, this issue connects to my general comment about the lack of scientific justification for the model modifications. If model parametrizations and data changes lead to almost negligible changes, the scientific rationale for implementing them should be discussed further.