Pith. sign in

REVIEW 5 major objections 5 minor 28 references

LD-RPMNet: Near-Sensor Diagnosis for Railway Point Machines

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LD-RPMNet classifies ten ZDJ9 railway point-machine states from switching sound at 98.86% accuracy while halving baseline parameters and FLOPs.

desk verdict The headline accuracy is arithmetically impossible on the stated 122-sample test set, and the architectural novelty is thin; the paper needs major corrections before it can be taken seriously. read the letter →

arxiv 2506.06346 v1 pith:225EDJLU submitted 2025-06-01 eess.SP cs.LG

Wei Li , Xiaochun Wu , Xiaoxi Hu , Yuxuan Zhang , Sebastian Bader , Yuhan Huang This is my paper
classification eess.SPcs.LG
keywords railwaypointmachinenear-sensordiagnosisfaultlightweightmodeldepthwiseseparableconvolutionbroadcastself-attentionacousticsignalanalysisZDJ9turnout
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

This paper is trying to establish that a railway point machine can be diagnosed where the sensor sits: on the sound of its switching motion, without shipping raw audio to the cloud. It proposes LD-RPMNet, a hybrid CNN-transformer classifier whose two modules, Multi-scale Depthwise Separable Convolution (MDSC) and Broadcast Self-Attention (BSA), are cut down enough for resource-constrained use. On 1,212 recordings of a ZDJ9 point machine spanning ten operating states, the model reports 98.86% test accuracy while reducing parameters and computational cost by roughly half relative to the CNN-transformer baseline. If the result holds, accurate acoustic fault diagnosis no longer needs a large server, which matters for real-time railway maintenance and for deployment on cheap, low-power trackside hardware.

What carries the argument

The central mechanism is the two lightweight modules that replace the expensive parts of a CNN-transformer hybrid. MDSC decomposes each cross-channel convolution into per-channel depthwise convolutions followed by a 1x1 pointwise convolution, and runs several kernel sizes in parallel so it keeps multi-scale local features while shrinking the parameter count. BSA replaces the query-key-value matrix multiplications and softmax exponentials of multi-head self-attention with broadcast operations that propagate global feature-weight information at far lower cost. Together, they are what let the model hold or improve accuracy while cutting the baseline's parameters and FLOPs by about half.

What would settle it

Evaluate LD-RPMNet and the CNN-transformer baseline with k-fold cross-validation or repeated random splits on the same 1,212 recordings. If the median accuracy gap between them is less than roughly two samples (about 1.6 percentage points) or changes sign, the claimed near-3% improvement is an artifact of a lucky split; additionally, deduplicating recordings from the same switching event would test whether the near-perfect scores rely on near-duplicate examples.

Watch

Extended reading notes

Core claim

The central claim is that a CNN-transformer model for point-machine fault diagnosis can be made much lighter without giving up accuracy. Replacing dense cross-channel convolutions with multi-scale depthwise separable convolution, and replacing multi-head self-attention with broadcast self-attention, simultaneously improves classification accuracy and roughly halves model size and computation. On the 122-sample test split, LD-RPMNet reports 98.86% accuracy with perfect precision, recall, and F1-score, using 0.48M parameters and 20.96M FLOPs with an inference time of 0.025 s. The paper presents this as evidence that near-sensor, sound-based fault diagnosis of ZDJ9 railway point machines is practical, since the model is small enough to reduce cloud dependence and improve real-time response.

Load-bearing premise

The load-bearing premise is that one fixed 7:2:1 split of the 1,212 sound samples into 122 test samples represents the true difficulty of the ten-class problem, so the reported 2.26-percentage-point accuracy gain over the CNN-transformer baseline is not just a two- or three-sample fluctuation.

Editorial extensions

If this is right

  • A trained LD-RPMNet can be embedded near the point machine and classify normal operation plus nine fault types from a switching sound in about 25 ms, enabling alarms without waiting for cloud inference.
  • The roughly 50% reduction in parameters and FLOPs relative to the CNN-transformer baseline means the same diagnosis could run on far cheaper, lower-power hardware.
  • Because raw sound no longer has to leave the trackside, bandwidth needs drop and operational data stays more private.
  • Other lightweight hybrid models such as CLFormer and Convformer NSE are smaller still but lose 4-8 percentage points of accuracy, suggesting the MDSC-plus-BSA combination hits a better operating point for this task.

Reading between the lines

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

  • A natural deployment step the paper does not run is porting the model to a microcontroller or edge processor and measuring energy per inference; the reported 0.025 s inference time comes from a desktop GPU, not from near-sensor hardware.
  • The same MDSC-plus-BSA recipe should transfer to other subtle acoustic monitoring tasks, such as bearing wear or contact wear, whenever the target device is battery- or resource-limited; that is a testable extension rather than a claim in the paper.
  • Because a full-quality ResNet18 beats LD-RPMNet by only 0.28 percentage points while using about eight times more parameters, the interesting engineering question left open is how to weigh a near-tie accuracy difference against an eight-fold resource saving in field deployment.
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

5 major / 5 minor

Summary. The paper proposes LD-RPMNet, a lightweight CNN-Transformer hybrid for sound-based fault diagnosis of ZDJ9 railway point machines. The architecture combines a Multi-scale Depthwise Separable Convolution (MDSC) module and a Broadcast Self-Attention (BSA) module. The authors report an accuracy of 98.86% on a 122-sample test set, along with approximately 50% reductions in parameters and computational cost relative to an in-house CNN-Transformer baseline (CNT). The paper includes an ablation study and comparisons with several lightweight and standard classifiers.

Significance. The application domain—near-sensor fault diagnosis for railway point machines—is practically relevant, and the design goal of a lightweight CNN-Transformer hybrid with reduced computational cost is well motivated. If the reported results were reproducible, the paper would offer a useful contribution to edge/embedded fault diagnosis. The authors provide a clear architecture description and figures, and their ablation design (MDSC and BSA added separately and together) is a reasonable way to isolate each module's contribution. However, the central quantitative claims are internally inconsistent: the reported accuracy values cannot be achieved on the stated test set, and the reported precision/recall/F1 values contradict the accuracy values. These inconsistencies mean that the paper's main evidence for near-sensor suitability cannot currently be assessed, and the contribution therefore does not meet the standard for publication.

major comments (5)
  1. [Section III-B, Table II, Table III] The test set is explicitly stated to contain 122 samples (7:2:1 split of 1,212 samples). On a discrete test set of 122 samples, achievable accuracy values must be k/122, i.e., 98.36% (120 correct) or 99.18% (121 correct). The reported values of 98.86% (LD-RPMNet), 96.60% (CNT), 97.50% (MobileNet), and 99.14% (ResNet18) are all arithmetically impossible. This directly invalidates the abstract's headline claim of 98.86% accuracy and calls into question every accuracy number in Tables II and III.
  2. [Table III] For LD-RPMNet, MobileNet, and ResNet18, the table reports accuracy below 100% while precision, recall, and F1-score are all 100%. In a multiclass confusion matrix, any misclassified sample—which the reported accuracy implies—would force at least one class's precision or recall below 100%. The metrics are therefore mutually contradictory, and the table cannot be used as evidence for the model's diagnostic performance.
  3. [Abstract, Section IV-B, Table III] The abstract claims that the optimized model 'reduces parameter count and computational complexity by 50%'. Table III shows CNT at 1.30M parameters and LD-RPMNet at 0.48M, a reduction of approximately 63%, while FLOPs drop from 39.41M to 20.96M, approximately 47%. The parameter reduction claim is not supported by the table. Additionally, the 'nearly 3% improvement' corresponds to a difference of 2.26 percentage points (98.86% vs. 96.60%), which on a 122-sample test set is about three samples; without repeated splits or error bars, this ranking is fragile.
  4. [Section IV-B, Reference [24]] The paper does not compare LD-RPMNet against LiConvFormer [24], which already combines separable multiscale convolution and broadcast self-attention—the two core components of LD-RPMNet. The reported improvements are calculated only against the in-house CNT baseline, not against the closest prior architecture. A fair comparative evaluation and a discussion of the differences relative to [24] are necessary to support the claim of novelty and the 'nearly 3%' improvement.
  5. [Section III-A, Section III-B] The dataset and experimental protocol are underdescribed: there is no information on how the 1,212 sound samples were segmented from continuous recordings, whether samples from the same switching event appear in multiple splits, or what preprocessing (filtering, normalization, etc.) was applied. The single 7:2:1 split yields only 122 test samples, and no cross-validation, repeated runs, or confidence intervals are reported. The reported differences between methods are therefore not statistically substantiated.
minor comments (5)
  1. [Table II] The table heading contains a typo: 'REAULTS' should be 'RESULTS'.
  2. [Section II-B] The text uses 'MSHA' when referring to multi-head self-attention; the standard abbreviation is 'MHSA'.
  3. [Section III-B] The 7:2:1 split of 1,212 samples gives fractional expected sizes of 848.4/242.4/121.2; the reported values 845/245/122 are inconsistent with the stated ratio unless rounding is explained. Please clarify.
  4. [Figures 7 and 8] The confusion matrices in Figures 7 and 8 are difficult to read in the provided resolution; axis labels and class names should be clearly visible, and the color scale should be specified.
  5. [References] Reference [24] is highly relevant and should be directly compared in the experiments; currently it is only cited as the source of the BSA mechanism.
Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced; the contributions are architectural combinations of existing modules. The free parameters listed are the main hand-chosen hyperparameters, several of which are not reported.

free parameters (4)
  • learning rate = 0.001
    Set by hand in Section III.B; not swept, and no sensitivity analysis is provided.
  • batch size = 16
    Chosen to balance speed and accuracy in Section III.B.
  • number of training epochs = 50
    Convergence observed at 40 epochs and training run to 50 epochs (Section III.B).
  • MDSC multi-scale kernel sizes = not reported
    The number and size of the parallel convolution kernels in Section II.A are never specified, though they directly control the model's accuracy and FLOP count.
assumptions (5)
  • domain assumption The ten fault classes in Table I are correctly labeled and mutually exclusive in the collected sound recordings.
    Dataset quality is load-bearing; label noise in a 1,212-sample private set would directly corrupt the 98.86% accuracy claim (Section III.A).
  • domain assumption The single 7:2:1 split of 1,212 samples into train, validation, and test is a reliable estimator of generalization.
    Section III.B defines the split (845/245/122) and the paper reports no cross-validation or repeated runs.
  • domain assumption Sound recordings captured in clear weather with no additional noise are representative of real operational conditions.
    Section III.A states acquisition was under clear weather without additional noise interference; no out-of-distribution test is performed.
  • standard math Depthwise separable convolution followed by pointwise convolution is a valid decomposition of cross-channel convolution.
    Section II.A relies on this standard decomposition (Eqs. 1-2).
  • domain assumption FLOPs and parameter counts are computed with a consistent methodology across all compared models.
    No measurement code is provided, so comparability of Table III cannot be checked.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LD-RPMNet: Near-Sensor Diagnosis for Railway Point Machines." pith.science (2026). https://pith.science/paper/225EDJLU

@misc{pith2026250606346,
  author       = {Pith},
  title        = {Pith review of: LD-RPMNet: Near-Sensor Diagnosis for Railway Point Machines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/225EDJLU}},
  note         = {Machine review of arXiv:2506.06346}
}
read the original abstract

Near-sensor diagnosis has become increasingly prevalent in industry. This study proposes a lightweight model named LD-RPMNet that integrates Transformers and Convolutional Neural Networks, leveraging both local and global feature extraction to optimize computational efficiency for a practical railway application. The LD-RPMNet introduces a Multi-scale Depthwise Separable Convolution (MDSC) module, which decomposes cross-channel convolutions into pointwise and depthwise convolutions while employing multi-scale kernels to enhance feature extraction. Meanwhile, a Broadcast Self-Attention (BSA) mechanism is incorporated to simplify complex matrix multiplications and improve computational efficiency. Experimental results based on collected sound signals during the operation of railway point machines demonstrate that the optimized model reduces parameter count and computational complexity by 50% while improving diagnostic accuracy by nearly 3%, ultimately achieving an accuracy of 98.86%. This demonstrates the possibility of near-sensor fault diagnosis applications in railway point machines.

Figures

Figures reproduced from arXiv: 2506.06346 by the authors.

Figure 1
Figure 1. Overview of the proposed near-sensor RPM fault diagnosis approach and its implementation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 4
Figure 4. The structure of BSA. transformations are applied to generate the query, key, and value matrices. For each attention head, the dot product between the query and key matrices is computed, followed by the application of the softmax function to derive the attention weights. These weights are then used to perform a weighted sum of the value matrix, yielding the output for each head. The outputs from multiple heads are s… view at source ↗
Figure 3
Figure 3. The structure of MDSC. Concat(·) represents the concatenation operation; C1 repre￾sents the input channel dimension, N1 denotes the input time dimension, C2 refers to the output channel dimension, and N2 indicates the output time dimension. Subsequently, the output signals from the depthwise con￾volution are concatenated, and a cross-channel convolution with a kernel size of 1 is employed to integrate information ac… view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: ZDJ9-type RPM and sound sensor location. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 7
Figure 7. Figure 7: Confusion matrix of ablation study: (a) CNT, (b) CNT+MDSC, (c) [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Confusion matrix of comparative experiment: (a) CNT, (b) LD [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 24 canonical work pages

  1. [24]

    Liconvformer: A lightweight fault diagnosis framework using separable multiscale con- volution and broadcast self-attention,

    S. Yan, H. Shao, J. Wang, X. Zheng, and B. Liu, “Liconvformer: A lightweight fault diagnosis framework using separable multiscale con- volution and broadcast self-attention,” Expert Systems with Applications , vol. 237, p. 121338, 2024

    work page 2024

  2. [1]

    A nonlinear dynamics method using multi-sensor signal fusion for fault diagnosis of rotating machinery,

    F. Chen, Z. Zhao, X. Hu, D. Liu, X. Yin, and J. Yang, “A nonlinear dynamics method using multi-sensor signal fusion for fault diagnosis of rotating machinery,” Advanced Engineering Informatics , vol. 65, p. 103190, 2025

    work page 2025

  3. [2]

    Tensor poincar´e plot index: A novel nonlinear dynamic method for extracting abnormal state information of pumped storage units,

    F. Chen, C. Ding, X. Hu, X. He, X. Yin, J. Yang, and Z. Zhao, “Tensor poincar´e plot index: A novel nonlinear dynamic method for extracting abnormal state information of pumped storage units,” Reliability Engi- neering & System Safety , vol. 254, p. 110607, 2025

    work page 2025

  4. [3]

    Railway automatic switch stationary contacts wear detection under few-shot occasions,

    X. Hu, Y . Cao, Y . Sun, and T. Tang, “Railway automatic switch stationary contacts wear detection under few-shot occasions,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 9, pp. 14 893–14 907, 2022

    work page 2022

  5. [4]

    Accurate modeling of railway turnout systems and high-precision state estimation methods,

    J. Liu, T. Wen, X. Fang, B. Cai, and C. Roberts, “Accurate modeling of railway turnout systems and high-precision state estimation methods,” IEEE Sensors Journal , 2025, doi: 10.1109/JSEN.2025.3545619

    arXiv 2025

  6. [5]

    Data-driven technology of fault diagnosis in railway point machines: Review and challenges,

    X. Hu, Y . Cao, T. Tang, and Y . Sun, “Data-driven technology of fault diagnosis in railway point machines: Review and challenges,” Transportation Safety and Environment , vol. 4, no. 4, p. tdac036, 2022

    work page 2022

  7. [6]

    A novel fault diagnosis model based on deep feature fusion network under imbalanced data: towards railway dual-switch machines traction occasion,

    S. Xiao, Q. Feng, X. Hu, Y . Song, G. Cong, Z. Yao, and H. Li, “A novel fault diagnosis model based on deep feature fusion network under imbalanced data: towards railway dual-switch machines traction occasion,” Measurement Science and Technology , vol. 36, no. 1, p. 016219, 2024

    work page 2024

  8. [7]

    Vibration signal-based fault diagnosis of railway point machines via double-scale cnn,

    X. Chen, X. Hu, T. Wen, and Y . Cao, “Vibration signal-based fault diagnosis of railway point machines via double-scale cnn,” Chinese Journal of Electronics , vol. 32, no. 5, pp. 972–981, 2023

    work page 2023

Show all 28 references
  1. [8]

    Fault detection for point machines: A review, challenges, and perspectives,

    X. Hu, T. Tang, L. Tan, and H. Zhang, “Fault detection for point machines: A review, challenges, and perspectives,” Actuators, vol. 12, no. 10, 2023

    2023

  2. [9]

    Tiny Machine Learn- ing for Damage Classification in Concrete Using Acoustic Emission Signals,

    V . Adın, Y . Zhang, B. Oelmann, and S. Bader, “Tiny Machine Learn- ing for Damage Classification in Concrete Using Acoustic Emission Signals,” in 2023 IEEE International Instrumentation and Measurement Technology Conference (I2MTC) , May 2023, pp. 1–6

    2023

  3. [10]

    Tiny-Machine-Learning-Based Sup- ply Canal Surface Condition Monitoring,

    C. Huang, X. Sun, and Y . Zhang, “Tiny-Machine-Learning-Based Sup- ply Canal Surface Condition Monitoring,” Sensors, vol. 24, no. 13, p. 4124, Jan. 2024

    2024

  4. [11]

    Railway switch machine fault diagnosis considering sensor abnormality scenarios,

    X. Hu, X. Zhang, Z. Wang, Y . Chen, J. Xia, Y . Du, and Y . Li, “Railway switch machine fault diagnosis considering sensor abnormality scenarios,” in 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) , 2023, pp. 4834–4839

    2023

  5. [12]

    Modified multi-scale symbolic dynamic entropy and fuzzy broad learning-based fast fault diagnosis of railway point machines,

    J. Liu, T. Wen, G. Xie, and Y . Cao, “Modified multi-scale symbolic dynamic entropy and fuzzy broad learning-based fast fault diagnosis of railway point machines,” Transportation Safety and Environment, vol. 5, no. 4, p. tdac065, 2023

    2023

  6. [13]

    Multi-time-scale variational mode decomposition-based robust fault diagnosis of railway point machines under multiple noises,

    J. Liu, T. Wen, G. Xie, Y . Cao, and C. Roberts, “Multi-time-scale variational mode decomposition-based robust fault diagnosis of railway point machines under multiple noises,” Chinese Journal of Electronics , vol. 33, no. 3, pp. 814–822, 2024

    2024

  7. [14]

    Simultaneous fault diagnosis for sensor and railway point machine for autonomous rail system,

    X. Hu, X. Zhang, F. Chen, Z. Liu, L. Tan, and T. Tang, “Simultaneous fault diagnosis for sensor and railway point machine for autonomous rail system,” in 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) , 2024, pp. 2132–2138

    2024

  8. [15]

    Leveraging Acoustic Emission and Machine Learning for Concrete Materials Damage Clas- sification on Embedded Devices,

    Y . Zhang, V . Adin, S. Bader, and B. Oelmann, “Leveraging Acoustic Emission and Machine Learning for Concrete Materials Damage Clas- sification on Embedded Devices,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–8, 2023

    2023

  9. [16]

    Compari- son of tiny machine learning techniques for embedded acoustic emission analysis,

    U. Muthumala, Y . Zhang, L. S. Martinez-Rau, and S. Bader, “Compari- son of tiny machine learning techniques for embedded acoustic emission analysis,” in 2024 IEEE 10th World F orum on Internet of Things (WF- IoT), 2024, pp. 444–449

    2024

  10. [17]

    Enabling Autonomous Structural Inspections with Tiny Machine Learning on UA Vs,

    Y . Zhang, L. S. Martinez-Rau, B. Oelmann, and S. Bader, “Enabling Autonomous Structural Inspections with Tiny Machine Learning on UA Vs,” in2024 IEEE Sensors Applications Symposium (SAS) , Jul. 2024, pp. 1–6

    2024

  11. [18]

    Survey of quantization techniques for on-device vision- based crack detection,

    Y . Zhang, L. S. Martinez-Rau, Q. N. Phuong Vu, B. Oelmann, and S. Bader, “Survey of quantization techniques for on-device vision- based crack detection,” in 2025 IEEE International Instrumentation and Measurement Technology Conference (I2MTC) , May 2025, pp. 1–6

    2025

  12. [19]

    A Lightweight Convolutional Neural Network Model for Concrete Damage Classification using Acoustic Emissions,

    Y . Zhang, S. Bader, and B. Oelmann, “A Lightweight Convolutional Neural Network Model for Concrete Damage Classification using Acoustic Emissions,” in 2022 IEEE Sensors Applications Symposium (SAS), Aug. 2022, pp. 1–6

    2022

  13. [20]

    Sepal: Sepsis alerts on low power wearables with digital biomarkers and on-device tiny machine learning,

    M. Giordano, K. Dheman, and M. Magno, “Sepal: Sepsis alerts on low power wearables with digital biomarkers and on-device tiny machine learning,” IEEE Sensors Journal , pp. 1–1, 2024

    2024

  14. [21]

    Msit: A cross-machine fault diagnosis model for machine-level cnc spindle motors,

    Y . He and W. Shen, “Msit: A cross-machine fault diagnosis model for machine-level cnc spindle motors,” IEEE Transactions on Reliability , 2023

    2023

  15. [22]

    Transformers in vision: A survey,

    S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, “Transformers in vision: A survey,” ACM computing surveys (CSUR) , vol. 54, no. 10s, pp. 1–41, 2022

    2022

  16. [23]

    Cfspt: A lightweight cross-machine model for compound fault diagnosis of machine-level motors,

    Y . He and W. Shen, “Cfspt: A lightweight cross-machine model for compound fault diagnosis of machine-level motors,” Information Fusion, vol. 111, p. 102490, 2024

    2024

  17. [25]

    Convformer-nse: A novel end-to-end gearbox fault diagnosis framework under heavy noise using joint global and local information,

    S. Han, H. Shao, J. Cheng, X. Yang, and B. Cai, “Convformer-nse: A novel end-to-end gearbox fault diagnosis framework under heavy noise using joint global and local information,” IEEE/ASME Transactions on Mechatronics, vol. 28, no. 1, pp. 340–349, 2022

    2022

  18. [26]

    Clformer: A lightweight transformer based on convolutional embed- ding and linear self-attention with strong robustness for bearing fault diagnosis under limited sample conditions,

    H. Fang, J. Deng, Y . Bai, B. Feng, S. Li, S. Shao, and D. Chen, “Clformer: A lightweight transformer based on convolutional embed- ding and linear self-attention with strong robustness for bearing fault diagnosis under limited sample conditions,” IEEE Transactions on Instrume...

    2021

  19. [27]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications,

    A. G. Howard, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,” arXiv preprint arXiv:1704.04861 , 2017

    2017 arXiv

  20. [28]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.