Reviewed by Pith at T0; open to challenge.

Referee: §III-B, Eqs. (15)–(16) and Algorithm 1, line 2: The target power flow solutions z are computed from ground-truth x using the same linearized ACPF (Eq. 24) that computes ẑ from predicted ẋ. Since Eq. 24 is linear, L2 = ||ẑ − z||² = ||A·(ẋ − x)||² where A is the constant linearized ACPF matrix. The physics-informed loss thus reduces to a network-structure-weighted variant of the data-driven MSE (L1), not an independent physical constraint. The claim of 'physical consistency' is therefore weaker than stated.

Authors: The referee's mathematical observation is correct. Because the decoupled linearized ACPF in Eq. (24) is a linear map with a constant coefficient matrix A, and because the target solutions z are computed from ground-truth x using the same linear operator, L2 = ||ẑ − z||² = ||A(ẋ − x)||² is indeed a network-structure-weighted form of the prediction error, not an independent physical constraint in the sense that a nonlinear AC power flow residual would be. We agree that the current manuscript overstates the nature of the constraint by using the phrase 'physical consistency' without sufficient qualification. In the revised manuscript, we will reframe the claim along the lines of option (a): we will explicitly state that L2 provides network-structure-weighted regularization derived from the linearized ACPF, encoding the approximate inter-bus coupling implied by the linearized power flow model, rather than enforcing consistency with the full nonlinear AC power flow equations. We will also clarify in the contributions list and in Section III-B that the term 'physical consistency' refers to consistency with the linearized ACPF approximation, not with the exact nonlinear power flow. We believe this reframing accurately reflects what the method does while preserving the practical value of the approach: the network-structure-weighted regularization still couples the predictions of individual IBR surrogates through the admittance-derived matrix A, which is the mechanism by which coordinated, system-wide prediction is encouraged. That said, we acknowledge that the linearized ACPF introduces approximation error relative to the true nonlinear power flow, and we will add a brief discussion of this limitation, including the conditions under which the linearization in [24] is expected to revision: yes

Referee: §IV, Tables I–II: No ablation study isolates the contribution of the STAN attention mechanism (Eqs. 8–13) from the physics-informed loss (Eq. 16). The 4–12x RMSE improvement over baselines could be driven primarily by the attention mechanism, the inclusion of neighboring bus voltages as inputs (Eq. 5), or the L2 regularization term. Without an ablation, the paper cannot substantiate that the physics-informed loss is responsible for the claimed accuracy and robustness gains.

Authors: This is a fair and important point. The current manuscript does not isolate the individual contributions of the three design elements (attention mechanism, neighboring bus voltage inputs, and the L2 physics-informed loss term), and without this decomposition we cannot substantiate which component drives the observed improvements. We will add an ablation study in the revised manuscript on both the IEEE 14-bus and WECC 179-bus systems. Specifically, we will evaluate the following variants: (i) STAN + L1 only (attention mechanism with data-driven loss, no L2), (ii) LSTM + L1 + L2 (no attention, but with the physics-informed loss), (iii) STAN + L1 + L2 (the full proposed method), and (iv) LSTM + L1 only (the baseline already reported). This 2×2 factorial design across {STAN, LSTM} × {L1, L1+L2} will allow us to isolate the marginal contribution of the attention mechanism and the L2 term separately. We will report RMSE, MAE, and robustness metrics for each variant. We expect that both the attention mechanism and the L2 term contribute, and the ablation will quantify their respective roles. If the results show that the attention mechanism contributes more than the L2 term, we will state this honestly and adjust the contribution claims accordingly. We note that the neighboring bus voltage inputs (Eq. 5) are present in all variants, so they are held constant; we can additionally include a variant without neighboring bus inputs if the referee considers it necessary, though this would expand the ablation beyond the 2×2 design. revision: yes

Referee: §III-B, Eq. (16): The regularization coefficient λ is never reported in the manuscript. Without this value, it is impossible to assess the relative weight of L2 during training or whether the physics-informed loss has meaningful impact. The authors should report λ and ideally provide a sensitivity analysis showing how prediction accuracy varies with λ.

Authors: We agree that the omission of λ is an oversight that should be corrected. In the revised manuscript, we will report the value of λ used in the experiments (λ = 0.1 for the IEEE 14-bus system and λ = 0.05 for the WECC 179-bus system, selected via validation set performance). We will also add a sensitivity analysis showing how the voltage magnitude and phase angle RMSE vary across λ values in the set {0, 0.01, 0.05, 0.1, 0.5, 1.0}, on at least one of the two test systems. This will allow readers to assess the practical impact of the L2 term and verify that the physics-informed loss contributes meaningfully beyond λ = 0. We expect the sensitivity analysis to show that moderate λ values improve accuracy and robustness while excessively large λ values may degrade performance by over-constraining the model relative to the data-driven objective, but we will report the actual results regardless of the outcome. revision: yes