Pith. sign in

REVIEW 4 major objections 4 minor 20 references

Neural Controlled Differential Equations for EMT-Level Surrogate Modeling of Grid-Forming Inverters

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a Neural CDE trained on 1 kHz sampled inverter currents can replace the detailed electromagnetic-transient model of a grid-forming inverter in simulation, preserving damping and dominant oscillatory modes while steppi

desk verdict Solid Neural CDE adaptation for inverter surrogates, but the EMT-level claim outruns 1 kHz training data—reframe or validate. read the letter →

arxiv 2607.16258 v1 pith:SUEEKECQ submitted 2026-06-28 cs.LG cs.AI

Jiagang Qu , Yong Tao , Dan Wang , Enyi Li , Jingjing Qi , Ding Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords NeuralcontrolleddifferentialequationGrid-forminginverterElectromagnetictransientsimulationSurrogatemodelingAffinecontroldecompositionMultiscaledynamicsLong-horizonstabilityDampingpreservation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that a Neural Controlled Differential Equation (Neural CDE) can serve as a physically consistent surrogate for a grid-forming inverter in electromagnetic transient (EMT) simulation. The proposed affine-control form splits the learned dynamics into an autonomous part, a slow control path, a fast control path, and a residual term, which the authors argue mirrors the inverter's hierarchical multi-rate control structure. Evaluated against EMT-generated trajectories, the surrogate reproduces transient grid-current responses, keeps the dominant damped oscillation mode and an effective damping score of 0.85, and stays bounded over long rollouts. If this holds, component-level EMT simulation could advance at 1 ms rather than 10 microseconds, potentially making thousand-converter stations tractable without knowing manufacturers' detailed control parameters. The paper itself notes that kilohertz-scale LC resonance is suppressed at the 1 kHz sampling rate, so the fidelity claim concerns closed-loop dynamics below about 500 Hz.

What carries the argument

The central object is the affine-control Neural CDE, a controlled differential equation in which a neural network specifies the state derivative with respect to a continuous control path rather than just time. Concretely, dx = fθ(x)dt + gθ^(1)(x)dX1 + gθ^(2)(x)dX2 + hθ(x,u)dt with dX1 = u dt and dX2 = du, separating slow accumulated control from fast control changes and adding a residual drift term. Supporting machinery includes a Jacobian spectral-radius regularizer built from the discrete-time flow Jacobian I + Δt Jx, an augmented-state lift, and a Hilbert-transform damping score (S_damp = 0.85) used as a physics-coherent evaluation metric.

What would settle it

Construct a detailed EMT simulation of the same ANPC converter with a disturbance that strongly excites the LC-filter resonance (e.g., a grid-voltage step or a load transient) and compare the 100 kHz reference waveform with the surrogate's output. If the reference shows significant oscillatory energy above 500 Hz that the surrogate cannot reproduce, while the paper claims EMT-level fidelity, the central claim is falsified. A spectral comparison of the two outputs over the first few milliseconds would settle it.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a Neural CDE can learn an effective continuous-time vector field for a dq0 grid current from sampled control inputs, and that this field, when integrated, reproduces the EMT reference's transient shapes, damping envelope, and dominant oscillatory behavior. The mechanism is the affine decomposition dx = f(x)dt + g1(x)dX1 + g2(x)dX2 + h(x,u)dt, where X1 accumulates the control signal and X2 is the control signal itself, so slow supervisory commands and fast inner-loop variations enter through separate channels. A Jacobian-based regularizer penalizes spectral radii above one for the discrete-time flow at EMT step sizes, which the authors credit for bounde

Load-bearing premise

The load-bearing premise is that the 1 kHz sampled, low-pass-filtered dq0 grid current contains the dynamics that matter, so the kilohertz-scale LC-filter resonance and other fast electromagnetic transients that are removed before training are not needed for the intended EMT studies.

Editorial extensions

If this is right

  • Detailed EMT models advance at 10 µs; the learned surrogate advances at 1 ms, so if accurate it removes roughly two orders of magnitude of integration steps from component-level simulation loops.
  • The surrogate is trained from trajectories alone, meaning systems with converters from different manufacturers could be modeled without proprietary control parameters.
  • Preserved damping and dominant oscillation modes make the surrogate usable for transient stability and multi-converter interaction studies over seconds-long horizons.
  • The dual slow/fast pathway gives an interpretable separation between supervisory (droop/VSM-like) and inner-current-loop dynamics, which can support validation and fault diagnosis.
  • Because the training data are band-limited to 1 kHz, the 'EMT-level' claim is restricted to closed-loop dynamics up to roughly 500 Hz; the paper explicitly notes the bare LC resonance is suppressed at this sampling rate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that a surrogate trained on 1 kHz data cannot be used for studies where kilohertz-scale electromagnetic transients are the object of interest; one concrete extension would be to train on a second fast path sampled at 100 kHz to cover the LC resonance.
  • The success of the dX2 = du term suggests that the derivative of the control signal carries load-bearing information, pointing to a general design rule for learned power-electronic dynamics: include rate-of-change of the reference as an explicit feature.
  • The affine dual-path architecture is not inverter-specific; the same decomposition could be applied to synchronous-generator or STATCOM surrogate models, creating a library of interchangeable learned components for large-scale EMT platforms.
  • A testable extension is to re-train the regularizer with a distribution of time steps rather than a single nominal Δt, to see whether the stability guarantee transfers to solvers with variable step sizes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a Neural CDE model for grid-forming inverter dynamics. The model augments a base Neural ODE with an affine-control decomposition separating autonomous dynamics (f_θ), slow control-path effects (g(1)_θ dX1), fast control-path effects (g(2)_θ dX2), and a residual term (h_θ), and adds a Jacobian-based stability regularizer. Training and evaluation use EMT-generated trajectories: raw 100 kHz waveforms are anti-alias filtered and decimated to 1 kHz, the state is the dq0 grid current, and the control vector consists of voltage amplitude, phase, and dq0 capacitor voltages. The paper reports ablation and sensitivity studies, a damping score of 0.85, bounded long-horizon rollouts, and concludes that the surrogate is 'physically consistent' and 'suitable for EMT-level simulation studies.'

Significance. If the claims were fully supported, the paper would offer a useful data-driven component model for accelerating EMT studies, with a principled continuous-time formulation and a thoughtful distinction between slow and fast control pathways. Strengths include the clear problem motivation, the explicit affine/dual-path architecture, the inclusion of ablation and sensitivity analyses, and the transparency about the 1 kHz sampling and partial observability. However, the central claim is broader than the evidence: the training data are band-limited to 500 Hz, and all reported evaluations operate on that band-limited signal, so the 'EMT-level' label is not justified without high-rate validation or a substantial rescoping of the claims.

major comments (4)
  1. [Sec. 5.3, 4.3, 5.6, Abstract/Conclusion] The training data are decimated from 100 kHz to 1 kHz after an anti-aliasing filter (Sec. 5.3), giving a Nyquist frequency of 500 Hz. The physical LC resonance (L1=180 uH, Cf=40 uF) is ~1.9 kHz, and Secs. 4.3 and 5.6 concede that this resonance is 'largely suppressed at the 1 kHz sampling rate.' The reported experiments, including Figs. 2-4 and the damping score, therefore validate only the band-limited closed-loop dynamics. The Abstract and Sec. 6 nevertheless state that the surrogate captures 'high-frequency electromagnetic responses' and is suitable for 'EMT-level simulation studies.' This is a load-bearing overclaim: the training signal cannot contain the fast transients that distinguish EMT from slower dynamic simulation. The authors should either validate against the raw 100 kHz EMT waveforms or revise all EMT-level claims and the title/abstract to describe a below-Nyquist closed-l
  2. [Sec. 5.3, Tables 3-5, Sec. 5.6] All numerical results are point estimates from a single 80/10/10 split with a fixed random seed and 9 test trajectories (Sec. 5.3). No error bars, confidence intervals, or repeated-seed results are reported. The headline improvements in MSE and stability are small (e.g., stability 0.8385 vs 0.8374 in Table 3), and without uncertainty quantification it is not possible to assess whether these differences are meaningful. The damping score S_damp=0.85 in Sec. 5.6 is also reported as a single scalar without stating how it was aggregated or what its spread is. The authors should provide multiple seeds or bootstrap intervals, or at least per-trajectory error distributions.
  3. [Sec. 4.2, Table 3] The Jacobian regularizer L_stab (Eq. 15) is introduced as a device to 'suppress unstable modes,' but the ablations in Tables 3-5 never remove this regularizer. Consequently, the stability improvements attributed to the Neural CDE formulation are not isolated from the effect of regularization. Add a variant without L_stab (or a sweep of its weight) to demonstrate its contribution.
  4. [Sec. 4.3, Eq. (17), Sec. 5.6] The damping score is computed via a Hilbert-envelope fit and reported only as 'S_damp = 0.85' (Sec. 5.6). The paper does not specify whether this is an average over the 9 test cases, how outliers are handled, or what the per-case variability is. Since this score is one of the two quantitative physics-coherence metrics, its aggregation and uncertainty must be reported for the claim to be assessable.
minor comments (4)
  1. [Tables 3 and 4] Table 3: '2.76810.8463' should read '2.7681 0.8463.' Table 4: 'in-256-256256-256-256-out' and 'in-256-256256-256-out' appear to be typographical errors in the layer-size notation.
  2. [Eq. (14)] The approximation D_x Φ = I + Δt J is a first-order Euler approximation, but training uses RK4. The mismatch should be stated explicitly, or the proxy should be derived for the RK4 update.
  3. [Sec. 3] The derivation via a first-order Taylor expansion plus a residual h_θ is more a justification than a constraint; because h_θ is a free neural network, the affine form is not actually enforced. The text could clarify that this is an architectural ansatz rather than a mathematical consequence.
  4. [Sec. 5.6] The stated speed-up from a 10 μs EMT step to a 1 ms surrogate step is presented as 'roughly two orders of magnitude fewer integration steps,' but no actual runtime or wall-clock comparison is reported. A simple step-count or wall-clock measurement would strengthen the deployment claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Neural CDE is trained on held-out trajectories, the evaluation metrics are not fitted parameters, and the bibliography contains no self-citations.

full rationale

The derivation chain is self-contained and non-circular. The Neural CDE formulation (Eqs. 7–12) is taken from the external literature on controlled differential equations (Kidger et al., Lyons et al.) and is not justified by the paper's own results. The affine-control decomposition is derived by a first-order Taylor expansion of a Neural ODE (Eq. 10), not by fitting a target quantity; the dual-path and residual modules are explicit architectural choices. The Jacobian regularizer (Eq. 15) is a training objective, while the stability score (Eq. 18) and damping score (Eq. 17) are evaluation metrics computed on held-out test trajectories against EMT references. There is no evidence that any test metric is used as a training loss or that any parameter is fitted to the reported test quantities; the damping score, in particular, is measured after training by comparing envelope decay rates. The reference list contains no self-citations: all cited works are external, including the foundational Neural ODE and Neural CDE papers. The only substantive limitation is a bandwidth/scope issue: the 1 kHz anti-aliased sampling (Sec. 5.3) suppresses the high-frequency LC resonance, which the paper explicitly acknowledges in Sec. 4.3 and Sec. 5.6. That is a validity or generalization concern about the 'EMT-level' claim, not a circular reduction to the model's inputs. The central predictions (transient trajectory match, damping, long-horizon boundedness) are not forced by construction, so no circularity score above zero is warranted.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The model rests on standard CDE theory plus a strong domain assumption: that a 1 kHz-filtered, reduced-state ODE can represent the EMT-relevant behavior. The paper introduces no new physical entities, but several hand-chosen hyperparameters (filter cutoff, regularization weight, etc.) are unreported, making the contribution slightly under-specified.

free parameters (5)
  • Per-unit base values = 245 A, 1000 V
    Used to normalize current and voltage; chosen by hand, affects the loss landscape but not the physical result.
  • Low-pass filter cutoff/decimation = unspecified
    Raw 100 kHz EMT data are filtered and decimated to 1 kHz; the cutoff frequency is not reported but determines which dynamics are visible to the model.
  • Training integration window = 26 samples
    MSE is minimized over 26-sample windows; length affects truncation bias and training stability.
  • Jacobian regularization weight = not reported
    L_stab in Eq. (15) requires a scalar weight; the value is never given, making the stability/accuracy trade-off impossible to assess.
  • RK4 training step size = not reported
    Selected via ACF/PSD/AEP analysis, but the resulting step size is not stated.
assumptions (6)
  • standard math Neural CDE solution theory and numerical ODE solvers converge for the learned vector fields.
    Assumes well-posedness of the controlled ODE and correctness of RK4/DOPRI5 integration.
  • domain assumption GFM inverter dynamics under symmetric operation can be represented by an ODE whose state is the dq0 grid current only.
    Section 2.1/Eq. (1) and Section 5.3; LC filter inductor/capacitor states are not part of the learned state.
  • domain assumption 1 kHz sampled, low-pass-filtered trajectories contain sufficient information to learn an effective continuous-time vector field; high-frequency LC resonance can be neglected.
    Section 4.3 states the LC resonance is suppressed at 1 kHz and the model learns effective closed-loop dynamics.
  • domain assumption DC-bus voltage is constant and the grid is balanced.
    Section 2.1 and Table 2; violations would change the dynamics.
  • ad hoc to paper First-order Taylor expansion of the control-dependent term, plus linear residual h_theta, captures the real control effects.
    Section 3, Eq. (10)-(11); higher-order terms are absorbed into a learned residual without justification that this is sufficient.
  • ad hoc to paper Partition of inputs into slow path (u dt) and fast path (du) matches the inverter's hierarchical control.
    Section 3, Eq. (12); claimed but not derived from the control hardware.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Controlled Differential Equations for EMT-Level Surrogate Modeling of Grid-Forming Inverters." pith.science (2026). https://pith.science/paper/SUEEKECQ

@misc{pith2026260716258,
  author       = {Pith},
  title        = {Pith review of: Neural Controlled Differential Equations for EMT-Level Surrogate Modeling of Grid-Forming Inverters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SUEEKECQ}},
  note         = {Machine review of arXiv:2607.16258}
}
read the original abstract

The application of artificial intelligence methods in power electronic converter modeling is becoming increasingly widespread, but existing applications still face many challenges, such as difficulties in multi-time-scale hybrid analysis and the lack of physics-aware evaluation criteria and constraints, resulting in poor performance. This paper proposes a Neural Controlled Differential Equation (Neural CDE) framework for learning continuous-time surrogate models of grid-forming inverters for electromagnetic transient (EMT) simulation, which relaxes the constraint of fixed sampling rates and enables multi-time-scale control analysis. Then, an affine-control formulation with dual slow/fast pathways is proposed to capture the hierarchical and multiscale behavior of converter dynamics, and a physics-inspired regularization method is utilized to enhance stability and coherence. Evaluated on EMT-generated trajectories, the model accurately reproduces transient responses, preserves effective damping and the dominant oscillatory characteristics, and maintains bounded long-horizon rollouts. The results show that Neural CDE-based component modeling offers a physically consistent surrogate modeling approach for EMT-level simulation studies.

Figures

Figures reproduced from arXiv: 2607.16258 by the authors.

Figure 1
Figure 1. Structure of Grid Forming Converter. from discretely sampled control signals. 2. Affine-control decomposition. The proposed affine form separates intrinsic system dynam￾ics and control-driven effects, aligning with the inverter’s hierarchical control structure and capturing multi-time-scale behaviors. 3. Dual-path control embedding. A dual control pathway, representing slow and fast variations of control inputs, is … view at source ↗
Figure 2
Figure 2. Comparison of grid current in the dq frame between EMT simulation and the Neural CDE model under various test cases. ∥xˆ(t) − x(t)∥ denote the instantaneous state error. Stability Score = 1 − 1 T − ∆t Z T −∆t 0 " log e(t + ∆t) + ε e(t) + ε # + dt (18) The integrand retains only the positive part of the local logarithmic error-growth rate, so the score equals one when the prediction error never amplifies along the ro… view at source ↗
Figure 3
Figure 3. Comparison of grid current in the abc frame between EMT simulation and the Neural CDE model under various test cases. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Long-horizon rollout of the Neural CDE model showing bounded state trajectories over [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 1 canonical work pages

  1. [1]

    Muntathir Al Talaq, Muhammad Bakr Abdelghany, Ahmed Al-Durra, Hatem Zeineldin, and Tarek EL-Fouly. A sophisticated grid-forming dispatchable-virtual oscillator control with ro- bust inner current control for improving transient stability.IEEE Transactions on Power Electronics, 40(10):15064–15079, 2025. doi: 10.1109/TPEL.2025.3575573

    arXiv 2025

  2. [2]

    Transient stability analysis of grid forming converters based on damping energy visualization and geometry approximation

    Jintao Lei, Xin Xiang, Boxin Liu, Wuhua Li, and Xiangning He. Transient stability analysis of grid forming converters based on damping energy visualization and geometry approximation. IEEE Transactions on Industrial Electronics, 71(3):2510–2521, 2024. doi: 10.1109/TIE.2023. 3269474

    work page doi:10.1109/tie.2023 2024

  3. [3]

    Transient stability analysis and control design of droop-controlled voltage source converters considering current limitation.IEEE Transactions on Smart Grid, 10(1):578–591, 2019

    Linbin Huang, Huanhai Xin, Zhen Wang, Leiqi Zhang, Kuayu Wu, and Jiabing Hu. Transient stability analysis and control design of droop-controlled voltage source converters considering current limitation.IEEE Transactions on Smart Grid, 10(1):578–591, 2019. doi: 10.1109/ TSG.2017.2749259

    arXiv 2019

  4. [4]

    McGraw-Hill, 1994

    Prabha Kundur.Power System Stability and Control. McGraw-Hill, 1994

    1994

  5. [5]

    Yitong Li, Yunjie Gu, Yue Zhu, Adri` a Junyent-Ferr´ e, Xin Xiang, and Timothy C. Green. Impedance circuit model of grid-forming inverter: Visualizing control algorithms as circuit elements.IEEE Transactions on Power Electronics, 36(3):3377–3395, 2021. doi: 10.1109/ TPEL.2020.3015158

    arXiv 2021

  6. [6]

    Graph neural solver for power systems

    Balthazar Donon, Benjamin Donnot, Isabelle Guyon, and Antoine Marot. Graph neural solver for power systems. In2019 international joint conference on neural networks (ijcnn), pages 1–8. IEEE, 2019

    2019

  7. [7]

    Foundation models for the electric power grid.Joule, 8(12):3245–3258, 2024

    Hendrik F Hamann, Blazhe Gjorgiev, Thomas Brunschwiler, Leonardo SA Martins, Al- ban Puech, Anna Varbella, Jonas Weiss, Juan Bernabe-Moreno, Alexandre Blondin Mass´ e, Seong Lok Choi, et al. Foundation models for the electric power grid.Joule, 8(12):3245–3258, 2024

    2024

  8. [8]

    Modeling power electronic converters using a method based on long-short term memory (lstm) networks

    Pouria Qashqai, Kamal Al-Haddad, and Rawad Zgheib. Modeling power electronic converters using a method based on long-short term memory (lstm) networks. InIECON 2020 The 46th Annual Conference of the IEEE Industrial Electronics Society, pages 4697–4702. IEEE, 2020

    2020

Show all 20 references
  1. [9]

    Physics-informed neural networks for phase locked loop transient stability assessment.Electric Power Systems Research, 236:110790, 2024

    Rahul Nellikkath, Ilgiz Murzakhanov, Spyros Chatzivasileiadis, Andreas Venzke, and Moham- mad Kazem Bakhshizadeh. Physics-informed neural networks for phase locked loop transient stability assessment.Electric Power Systems Research, 236:110790, 2024. 15

    2024

  2. [10]

    Neural network models and transfer learning for impedance modeling of grid-tied inverters

    Yufei Li, Yicheng Liao, Xiongfei Wang, Lars Nordstr¨ om, Prateek Mittal, Minjie Chen, and H Vincent Poor. Neural network models and transfer learning for impedance modeling of grid-tied inverters. In2022 IEEE 13th International Symposium on Power Electronics for Distributed Ge...

    2022

  3. [11]

    Feasibility study of neural ode and dae modules for power system dynamic component modeling.IEEE Transac- tions on Power Systems, 38(3):2666–2678, 2022

    Tannan Xiao, Ying Chen, Shaowei Huang, Tirui He, and Huizhe Guan. Feasibility study of neural ode and dae modules for power system dynamic component modeling.IEEE Transac- tions on Power Systems, 38(3):2666–2678, 2022

    2022

  4. [12]

    Christian Moya, Guang Lin, Tianqiao Zhao, and Meng Yue. On approximating the dynamic re- sponse of synchronous generators via operator learning: A step towards building deep operator- based power grid simulators.arXiv preprint arXiv:2301.12538, 2023

    2023 arXiv

  5. [13]

    Data-driven modeling of grid-forming inverter dynamics using power hardware-in-the-loop experimentation.IEEE Access, 12:52267–52281, 2024

    Nischal Guruwacharya, Soham Chakraborty, Govind Saraswat, Richard Bryce, Timothy M Hansen, and Reinaldo Tonkoski. Data-driven modeling of grid-forming inverter dynamics using power hardware-in-the-loop experimentation.IEEE Access, 12:52267–52281, 2024

    2024

  6. [14]

    Neuro-modeling infused emt analytics.IEEE Transactions on Consumer Electronics, 2025

    Qing Shen, Yifan Zhou, Peng Zhang, Yacov A Shamash, Roshan Sharma, and Bo Chen. Neuro-modeling infused emt analytics.IEEE Transactions on Consumer Electronics, 2025

    2025

  7. [15]

    Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    2018

  8. [16]

    Snode: Spectral dis- cretization of neural odes for system identification.arXiv preprint arXiv:1906.07038, 2019

    Alessio Quaglino, Marco Gallieri, Jonathan Masci, and Jan Koutn ´ ık. Snode: Spectral dis- cretization of neural odes for system identification.arXiv preprint arXiv:1906.07038, 2019

    1906 arXiv

  9. [17]

    Neural controlled differential equations for irregular time series.Advances in neural information processing systems, 33: 6696–6707, 2020

    Patrick Kidger, James Morrill, James Foster, and Terry Lyons. Neural controlled differential equations for irregular time series.Advances in neural information processing systems, 33: 6696–6707, 2020

    2020

  10. [18]

    Springer, 2007

    Terry J Lyons, Michael Caruana, and Thierry L´ evy.Differential equations driven by rough paths: Ecole d’Et´ e de Probabilit´ es de Saint-Flour XXXIV-2004. Springer, 2007

    2004

  11. [19]

    Augmented neural odes.Advances in neural information processing systems, 32, 2019

    Emilien Dupont, Arnaud Doucet, and Yee Whye Teh. Augmented neural odes.Advances in neural information processing systems, 32, 2019

    2019

  12. [20]

    How to train your neural ode: the world of jacobian and kinetic regularization

    Chris Finlay, J¨ orn-Henrik Jacobsen, Levon Nurbekyan, and Adam Oberman. How to train your neural ode: the world of jacobian and kinetic regularization. InInternational conference on machine learning, pages 3154–3164. PMLR, 2020. 16

    2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.