Made crvernon paper fixes

JOSS/paper.md

@@ -58,7 +58,7 @@ U_pred = esn.forecast(fcast_len=1000, res_state=R[-1])
 
 Built-in visualization tools make it easy to evaluate model performance. Varying the hyperparameters of built-in models lets users explore how RC performance depends on configuration choices. ORC's JAX foundation makes scaling to higher-dimensional parallel reservoir architectures equally simple.
 
-Much RC research is aimed at designing performant reservoir architectures. ORC makes this easy through its use of abstract base classes for Embedding, Driver, and Readout layers. Users need only define forward pass logic to integrate a new reservoir topology or readout strategy, while reusing the rest of the framework. This modular approach also enables ablation studies on how different components affect RC performance. Because of ORC's functional approach in JAX, built-in and user-created models provide end-to-end differentiability by default. This enables gradient-based optimization of input sequences for control problems. This also makes ORC well suited to integrate with deep learning models such as those presented in [@ozalp2023reconstruction; @ozalp2025real].
+Much RC research is aimed at designing performant reservoir architectures. ORC makes this easy through its use of abstract base classes for Embedding, Driver, and Readout layers. Users need only define forward pass logic to integrate a new reservoir topology or readout strategy, while reusing the rest of the framework. This modular approach also enables ablation studies on how different components affect RC performance. Because of ORC's functional approach in JAX, built-in and user-created models provide end-to-end differentiability by default. This enables gradient-based optimization of input sequences for control problems. This also makes ORC well suited to integrate with deep learning models such as those presented by Özalp et al. [-@ozalp2023reconstruction; -@ozalp2025real].
 
 # State of the Field
 | | **ORC** | **ReservoirPy** | **RC.jl** |
@@ -81,7 +81,7 @@ First, ORC was *designed* on top of JAX [@jax2018github] and Equinox [@kidger202
 Second, ORC has a different built-in feature set. ORC supports continuous-time reservoir dynamics via Diffrax [@kidger2021on], allowing users to define reservoir equations as ordinary differential equations solved with adaptive-step integrators. ORC also supports novel architectures such as Taylor-expanded and GRU-based drivers alongside standard echo state networks. ReservoirPy does not provide these functionalities. 
 Third, ORC models compose seamlessly with other deep learning models in Equinox, whereas ReservoirPy is suited to standalone RC tasks. 
 
-Other open source libraries for RC include Pytorch-ESN [@nardo2018pytorchesn] and ReservoirComputing.jl [@martinuzzi2022reservoircomputing]. Pytorch-ESN allows for integration with other PyTorch models, but is not as widely adopted and offers a much more limited feature set than ORC or ReservoirPy. For instance, it does not support sparse matrix operations or parallel reservoirs. ReservoirComputing.jl is a widely used Julia library for RC that influenced many of ORC's design choices, but lacks Python ML ecosystem integration. Table 1 summarizes the functionality of ORC, ReservoirPy, and ReservoirComputing.jl. 
+Other open source libraries for RC include PyTorch-ESN [@nardo2018pytorchesn] and ReservoirComputing.jl [@martinuzzi2022reservoircomputing]. PyTorch-ESN allows for integration with other PyTorch models, but is not as widely adopted and offers a much more limited feature set than ORC or ReservoirPy. For instance, it does not support sparse matrix operations or parallel reservoirs. ReservoirComputing.jl is a widely used Julia library for RC that influenced many of ORC's design choices, but lacks Python ML ecosystem integration. Table 1 summarizes the functionality of ORC, ReservoirPy, and ReservoirComputing.jl. 
 
 
 # Software Design
@@ -104,12 +104,12 @@ ORC also includes a data generation module with ODE and PDE integrators for stan
 
 ![ORC with GPU acceleration enables significantly faster performance than ReservoirPy, even when using ReservoirPy's JAX backend. Panel (a) shows the time per forecast step of RC models with varying reservoir dimension trained to forecast the Lorenz system, while panel (b) shows training time for RC models with fixed reservoir dimension of 2000 but a varying number of training samples. Performance of the two libraries with and without GPU acceleration are shown. GPU results were obtained on an NVIDIA A40 GPU and CPU results were obtained with an Apple M2 chip. \label{fig:comp}](../imgs/benchmark_results_combined.png) 
 
-ORC enables easy reimplementation of architectures that integrate RC with larger neural networks [@ozalp2023reconstruction; @ozalp2025real], and makes it easier to iterate on these ideas than any existing library. We also benchmark ORC against ReservoirPy across reservoir sizes and the number of training samples \autoref{fig:comp}. We find that with GPU acceleration, ORC scales far more favorably than ReservoirPy. Moreover, ORC is the only current package that supports training RC for control tasks, enabled by its efficiency and the end-to-end differentiability that JAX and Equinox provide. RC-based control has recently been shown to be advantageous over other forms of RNN-based MPC [@williams2024reservoir]. ORC's performance also makes it well suited for the parallel architectures needed to extend RC to higher-dimensional settings.
+ORC enables easy reimplementation of architectures that integrate RC with larger neural networks [@ozalp2023reconstruction; @ozalp2025real], and makes it easier to iterate on these ideas than any existing library. We also benchmark ORC against ReservoirPy across reservoir sizes and numbers of training samples, as shown in \autoref{fig:comp}. We find that with GPU acceleration, ORC scales far more favorably than ReservoirPy. Moreover, ORC is the only current package that supports training RC for control tasks, enabled by its efficiency and the end-to-end differentiability that JAX and Equinox provide. RC-based control has recently been shown to be advantageous over other forms of RNN-based MPC [@williams2024reservoir]. ORC's performance also makes it well suited for the parallel architectures needed to extend RC to higher-dimensional settings.
 
 
 # AI Usage Disclosure
 
-Claude (Anthropic) was used for code assistance during code debugging, proofreading this paper, and the generation of the architecture visualization SVG. ChatGPT (OpenAI) was used to generate the ORC logo. All generated code was reviewed, tested, and validated by the authors.
+Claude (Anthropic, Opus 4-4.8) was used for code assistance during debugging, for proofreading and editing the text of this paper, and for generating the architecture visualization SVG (\autoref{fig:architecture}). ChatGPT (OpenAI, GPT-4o) was used to generate the ORC logo. All AI-assisted outputs, including code, paper text, and figures, were reviewed, edited, and validated by the authors before inclusion.
 
 # Acknowledgements
 The authors acknowledge support from the NSF AI Institute in Dynamic Systems (grant number 2112085). The authors also thank Anastasia Bizyaeva, Noa Kaplan, and Ling-Wei Kong for insightful conversations.