Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Efficient Training of Deep Networks using Guided Spectral Data Selection: A Step Toward Learning What You Need

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read GSTDS claims that selecting training data by the Fiedler vector of a batch's similarity graph, guided by a frozen reference model and a rising filter schedule, cuts compute up to fourfold while preserving or improving test accuracy.

desk verdict Fiedler-vector batch selection with a curriculum schedule is a novel combination, but the FLOPs claim omits the reference model's cost and the evaluation lacks baselines and seeds, so the main result is unsupported as written. read the letter →

arxiv 2507.04269 v1 pith:TELD2U7K submitted 2025-07-06 cs.CV cs.LG

classification cs.CVcs.LG
keywords datacurationspectralanalysisFiedlervectoractivelearningselectioncurriculumtrainingefficiencyimageclassification
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

GSTDS is a batch-level data curation method that trains a model on only a scheduled fraction of each batch, using a pre-trained reference model's features and a Fiedler-vector score to keep what it claims are the most informative images. The paper argues this achieves up to roughly four times fewer floating-point operations than standard training while matching or improving test accuracy, and that it beats the recent JEST data-curation baseline on small image benchmarks. The motivation is that most training compute is spent on redundant or uninformative samples; if spectral ranking can identify the useful ones, resource-efficient deep learning becomes a data-selection problem rather than a scaling problem. The method appears to help most on fine-grained datasets with many classes and few examples per class.

What carries the argument

The load-bearing object is the Fiedler vector: for each batch, the paper builds a cosine-similarity matrix from reference-model feature vectors, forms the graph Laplacian, and takes the eigenvector of the second-smallest eigenvalue. Ranking data by the size of its components is the claimed measure of each sample's significance to the batch's geometric structure. A sigmoid schedule sets the per-batch filter ratio (from about 0.18 to 0.88, averaging 0.3 over training), and the selected subset is half deterministic top-Fiedler components plus half random samples weighted by inverse reference-model loss.

What would settle it

Run GSTDS with the same filter-ratio schedule but replace the Fiedler top-half selection with random selection from each batch; if test accuracy stays about the same, the spectral ranking is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that ordering the samples in a batch by the Fiedler vector of their cosine-similarity graph, then keeping the top half and drawing the other half by inverse-loss-weighted random sampling, preserves or improves accuracy while cutting compute by up to four times. On Oxford-IIIT Pet, GSTDS reaches 89.69% test accuracy against 89.35% for standard training while using 8.18e13 FLOPs versus 30.1e13; on Oxford-Flowers it reaches 58.13% versus 41.14% at 2.12e13 versus 8.34e13 FLOPs. The paper interprets the scheduled, increasing filter ratio as an implicit curriculum that starts with a small informative subset and gradually admits more data.

Load-bearing premise

A batch's largest positive Fiedler-vector components mark its most informative points for training, so that keeping them and dropping the rest preserves learning.

Editorial extensions

If this is right

  • Training compute can be cut by up to four times without losing test accuracy on standard image benchmarks.
  • Under a fixed FLOP budget, GSTDS reaches higher accuracy than both standard training and JEST, especially on fine-grained datasets like Oxford-Flowers.
  • The benefit grows with task difficulty: the largest accuracy gains appear on many-class, few-samples-per-class datasets.
  • The time-variant filter schedule acts as an implicit curriculum, and smoother schedules (sigmoid) outperform oscillatory or AIMD-style schedules.
  • GSTDS continues improving at the 25-epoch cutoff on CIFAR-10 where baselines plateau, suggesting the savings could compound with longer training.

Reading between the lines

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

  • The paper never tests whether the Fiedler ordering beats random ordering at the same filter schedule; if it does not, the entire gain might come from the curriculum schedule and the reference-model losses rather than spectral geometry.
  • The text and the pseudocode assign different weights to the random-sampling half (inverse loss vs. normalized Fiedler components), so the actual mechanism is ambiguous and an implementation must choose one; an ablation between the two would settle which is active.
  • The same selection rule could be tested outside vision, e.g., on language-model pretraining tokens or multimodal batches, where joint example selection is already known to help; a spectral formulation would need to handle much larger batch graphs.
  • Because the reference model's features are frozen, GSTDS's gains may track how well the reference model's geometry matches the target task; using a less related reference model is a cheap, testable way to probe where the signal comes from.
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 / 6 minor

Summary. The paper proposes GSTDS, a data-selection method that uses a frozen pre-trained ResNet-50 to extract features, computes a per-batch cosine-similarity Laplacian, and retains a scheduled fraction of each batch by ranking Fiedler-vector components. The retained fraction is split between the top-ranked points and a weighted random sample, with a sigmoid schedule that increases the filter ratio over training. Experiments on Oxford-IIIT Pet, Oxford-Flowers, and CIFAR-10 compare GSTDS against standard full-data training and a modified JEST baseline, reporting test accuracies and FLOPs. The central claim is that GSTDS reduces compute by up to 4x 'without compromising performance.'

Significance. If the efficiency and accuracy claims held under a complete accounting of compute, GSTDS would be a practically useful contribution to data-curation and resource-efficient training. The paper has several genuine strengths: it evaluates on three standard image benchmarks, includes a JEST baseline, studies several filter-ratio schedules, and releases code. It is also an empirical, falsifiable method paper that does not rely on fitted predictions of the reported accuracies. However, the current evidence is not strong enough to establish the central claims, because the FLOPs accounting is incomplete, there is no random-subset control, and all comparisons are single runs without uncertainty estimates.

major comments (4)
  1. [§3.1 and Tables 2–4] The FLOPs numbers in Tables 2–4 cannot support the abstract's 'up to four times' claim because they omit mandatory components of the GSTDS pipeline. Section 3.1 requires a full reference-model forward pass over the entire training set to extract features; for Oxford-Flowers, a ResNet-50 pass over its 8,189 training images is on the order of 3×10^13 FLOPs under standard accounting, which already exceeds the 2.12×10^13 reported as the total GSTDS cost in Table 4. Including this overhead reduces the reported Flowers speedup from roughly 3.9x to about 1.5x. The paper also never defines what is included in 'Total FLOPs', and it is unclear whether the point-wise reference-model losses described in Section 3.3 require an additional forward pass or a target-class classifier; Algorithm 1's line 12 instead uses Fiedler-vector magnitudes. Until the FLOPs definition includes all required computation, the central efficiency claim is unverified and, for at least one dataset, likely overstated by a factor of roughly two or more.
  2. [§4.2] There is no control experiment that trains on randomly selected data with the same filter-ratio schedule. Because the schedule in Section 3.2 constrains the average data usage to 0.30, a random 30% subset per batch is the minimal baseline needed to attribute the results to Fiedler-vector selection rather than to the reduced data volume or the curriculum schedule. Without such a baseline, the accuracy differences in Tables 2–4 could in principle be reproduced by random subsampling, and the paper's central mechanism would be unsupported.
  3. [Tables 2–4 and §4.2] All reported comparisons appear to be single training runs, yet Section 4.2 describes the CIFAR-10 result as 'statistically comparable' despite a 0.67-point test-accuracy deficit (82.56 for standard training vs 81.89 for GSTDS in Table 3). With no error bars, standard deviations, or significance testing, the claim that GSTDS achieves 'notable reductions in computational requirements ... without compromising performance' (abstract) is not empirically supported. This is load-bearing because the efficiency claim is explicitly conditional on matching accuracy.
  4. [§3.3] The claim that ranking batch elements by the largest Fiedler-vector components identifies 'the most informative samples' is a heuristic with no theoretical derivation or empirical isolation in the paper. No ablation compares the deterministic top-half selection against random selection, selection by absolute Fiedler magnitude, or selection by reference-model loss; the signed ranking is particularly unexamined. Since the Fiedler ranking is the core differentiator of the method, the current experiments do not establish that spectral selection, rather than the schedule or the reduced data budget, is responsible for the reported outcomes.
minor comments (6)
  1. [Algorithm 1 and §3.3] Algorithm 1's line 12 defines sampling weights as normalized Fiedler-vector magnitudes, while the text in Section 3.3 defines weights as 1/(ℓ(q_j)+ε) using reference-model point-wise losses. This inconsistency makes it unclear which selection procedure was actually executed and how the 'exploration' component is computed.
  2. [§3.2] The schedule parameters a, b, k, and x0 are said to be 'simply optimized' to satisfy the stated integral and boundary constraints, but the exact parameter values and the optimization procedure are not reported; these are needed for reproducibility.
  3. [Table 3] The table title reads 'CIF AR-10 Pet over 25 epochs', which appears to be a copy-paste error; the dataset is CIFAR-10, not a Pet dataset.
  4. [§4.4] The SVM-region analysis in Figure 8 is not reproducible because the paper does not specify how the SVM was trained, what features it used, or how regions were assigned to decision boundaries.
  5. [§4.5 and Figure 10] The conclusion that label 94 is prioritized because of its higher intra-class diversity is purely observational; no quantitative measure of intra-class diversity or difficulty is provided, and the selection-frequency difference between two classes is reported without statistical context.
  6. [Figure 7 and Table 5] The text says GSTDS achieves '≈ 87.50%' accuracy at fixed FLOPs, but the corresponding Table 5 reports GSTDS test accuracies of 89.64%, 71.34%, and 80.42% across the three datasets; this number is unexplained.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper's accuracy and FLOPs results are external measurements, not predictions derived from fitted inputs or self-citations.

full rationale

GSTDS is an empirical method paper. The reported accuracies per dataset are measured outcomes of a training procedure, and the FLOPs numbers are counts of learner operations; neither quantity is derived from the parameters that were set in advance. The scheduled filter-ratio is fitted to explicit constraints (average 0.3, max 0.88, min 0.18) that define data usage, not to the reported test accuracies, so the efficiency results are not forced by construction. The Fiedler-vector ranking is introduced as a heuristic for informativeness, and the claim that it selects 'most informative' samples is an assumption being tested empirically, not a self-definitional derivation. The method's central comparison against JEST and standard training rests on external benchmark measurements; the ImageNet-pretrained reference model overlap with benchmark classes is a potential leakage or correctness concern, but it is not circularity. There are no load-bearing self-citations: the paper cites prior work by other groups for spectral analysis, batch selection, and curriculum learning, and none of those citations carries the main claim. One internal inconsistency exists (Section 3.3 defines weighted sampling via reference-model losses, while Algorithm 1 line 12 uses Fiedler-vector magnitudes), but an inconsistency is a reproducibility defect, not a circular reduction. The FLOPs accounting omitting the one-time ResNet-50 feature extraction pass is likewise a correctness objection to the speedup claim, not a case of a prediction being equivalent to its input. Therefore no circular step can be exhibited, and the honest score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The core machinery rests on two heuristics (Fiedler-vector ranking and reference-model loss weighting) plus a fitted schedule. The schedule parameters are only partially reported, and the text and pseudocode disagree on the sampling weights. No invented entities are introduced.

free parameters (2)
  • sigmoid schedule parameters (a, b, k, x0) = a=0.18, b=0.88; k and x0 not reported
    Section 3.2 defines the schedule f(x) and states the parameters are 'optimized' to satisfy average 0.3, min 0.18, max 0.88. k and x0 are not given, so the exact schedule cannot be reproduced.
  • epsilon in weight formula = not reported
    Section 3.3 defines w_j = 1/(loss(q_j)+epsilon); epsilon affects the sampling distribution and is not specified. Note Algorithm 1 does not use this formula.
assumptions (5)
  • domain assumption Fiedler vector component values rank data points by informativeness within a batch
    Section 3.3 states this as fact; no proof or empirical validation is provided.
  • domain assumption Cosine similarity between reference-model features is a valid measure of inter-sample geometry for selection
    Section 3.3 builds the Laplacian from cosine similarity; this choice is not compared with other similarity measures.
  • domain assumption Reference model point-wise loss is an informative signal for sampling
    Section 3.3 text uses inverse loss as sampling weight; the rationale is asserted but not evaluated. Algorithm 1 instead uses Fiedler magnitudes, adding ambiguity.
  • domain assumption Pretrained ImageNet ResNet-50 features transfer to CIFAR-10, Flowers, and Pets
    Section 3.1; the reference model is frozen and its features drive all selection, with no domain-adaptation step.
  • ad hoc to paper Single-run training comparisons are sufficient
    All tables report one value per metric; no seeds or variance are given, yet the paper draws comparative conclusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Training of Deep Networks using Guided Spectral Data Selection: A Step Toward Learning What You Need." pith.science (2026). https://pith.science/paper/TELD2U7K

@misc{pith2026250704269,
  author       = {Pith},
  title        = {Pith review of: Efficient Training of Deep Networks using Guided Spectral Data Selection: A Step Toward Learning What You Need},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TELD2U7K}},
  note         = {Machine review of arXiv:2507.04269}
}
read the original abstract

Effective data curation is essential for optimizing neural network training. In this paper, we present the Guided Spectrally Tuned Data Selection (GSTDS) algorithm, which dynamically adjusts the subset of data points used for training using an off-the-shelf pre-trained reference model. Based on a pre-scheduled filtering ratio, GSTDS effectively reduces the number of data points processed per batch. The proposed method ensures an efficient selection of the most informative data points for training while avoiding redundant or less beneficial computations. Preserving data points in each batch is performed based on spectral analysis. A Fiedler vector-based scoring mechanism removes the filtered portion of the batch, lightening the resource requirements of the learning. The proposed data selection approach not only streamlines the training process but also promotes improved generalization and accuracy. Extensive experiments on standard image classification benchmarks, including CIFAR-10, Oxford-IIIT Pet, and Oxford-Flowers, demonstrate that GSTDS outperforms standard training scenarios and JEST, a recent state-of-the-art data curation method, on several key factors. It is shown that GSTDS achieves notable reductions in computational requirements, up to four times, without compromising performance. GSTDS exhibits a considerable growth in terms of accuracy under the limited computational resource usage, in contrast to other methodologies. These promising results underscore the potential of spectral-based data selection as a scalable solution for resource-efficient deep learning and motivate further exploration into adaptive data curation strategies. You can find the code at https://github.com/rezasharifi82/GSTDS.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 21 canonical work pages

  1. [1]

    In: Pro- ceedings of the 26th Annual International Conference on Machine Learning, pp

    Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Pro- ceedings of the 26th Annual International Conference on Machine Learning, pp. 41–48 (2009)

  2. [2]

    Advances in neural information processing systems 33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakan- tan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901 (2020)

  3. [3]

    Neural computation 15(6), 1373–1396 (2003)

    Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data representation. Neural computation 15(6), 1373–1396 (2003)

  4. [4]

    Journal of artificial intelligence research 4, 129–145 (1996)

    Cohn, D.A., Ghahramani, Z., Jordan, M.I.: Active learning with statistical models. Journal of artificial intelligence research 4, 129–145 (1996)

  5. [5]

    Computer Networks and ISDN Systems 17(1), 1–14 (1989)

    Chiu, D.-M., Jain, R.: Analysis of the increase and decrease algorithms for congestion avoidance in computer networks. Computer Networks and ISDN Systems 17(1), 1–14 (1989)

  6. [6]

    Journal of Machine Learning Research 19(27), 1–21 (2018)

    Csiba, D., Richt´ arik, P.: Importance sampling for minibatches. Journal of Machine Learning Research 19(27), 1–21 (2018)

  7. [7]

    arXiv preprint arXiv:1906.11829 (2019)

    Coleman, C., Yeh, C., Mussmann, S., Mirzasoleiman, B., Bailis, P., Liang, P., Leskovec, J., Zaharia, M.: Selection via proxy: Efficient data selection for deep learning. arXiv preprint arXiv:1906.11829 (2019)

  8. [8]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., Fei-Fei, L.: ImageNet: A Large- Scale Hierarchical Image Database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255 (2009). https://doi.org/10.1109/CVPR.2009. 5206848

Show all 35 references
  1. [9]

    arXiv preprint arXiv:2406.17711 (2024)

    Evans, T., Parthasarathy, N., Merzic, H., Henaff, O.J.: Data curation via joint example selection further accelerates multimodal learning. arXiv preprint arXiv:2406.17711 (2024)

  2. [10]

    MIT Press, Cambridge, MA, USA (2016)

    Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press, Cambridge, MA, USA (2016)

  3. [11]

    In: International Conference on Machine Learning, pp

    Gal, Y., Islam, R., Ghahramani, Z.: Deep bayesian active learning with image data. In: International Conference on Machine Learning, pp. 1183–1192 (2017). PMLR

  4. [12]

    In: ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp

    Gao, Z., Koppel, A., Ribeiro, A.: Balancing rates and variance via adaptive batch-sizes in first-order stochastic optimization. In: ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5385–5389 (2020). https://doi.org/10.11...

  5. [13]

    In: 17 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 17 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778 (2016). https://doi.org/10.1109/CVPR.2016.90

  6. [14]

    In: International Conference on Machine Learning, pp

    Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning, pp. 4904–4916 (2021). PMLR

  7. [15]

    Technical report, University of Toronto (2009)

    Krizhevsky, A., Hinton, G.E., Nair, V.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)

  8. [16]

    arXiv preprint arXiv:2001.08361 (2020)

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

  9. [17]

    nature 521(7553), 436–444 (2015)

    LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436–444 (2015)

  10. [18]

    In: Aggarwal, C.C., Reddy, C.K

    Liu, J., Han, J.: Spectral clustering. In: Aggarwal, C.C., Reddy, C.K. (eds.) Data Clustering: Algorithms and Applications, pp. 177–200. Chapman and Hall/CRC, New York (2018). https://doi.org/10.1201/9781315373515

  11. [19]

    arXiv preprint arXiv:2402.11215 (2024)

    Lau, T.T.-K., Liu, H., Kolar, M.: Adadagrad: Adaptive batch size schemes for adaptive gradient methods. arXiv preprint arXiv:2402.11215 (2024)

  12. [20]

    arXiv preprint arXiv:1003.3689, 146 (2010)

    Manguoglu, M.: A highly efficient parallel algorithm for computing the fiedler vector. arXiv preprint arXiv:1003.3689, 146 (2010)

  13. [21]

    In: Pro- ceedings of the 18th ACM International Conference on Multimedia, pp

    Marcel, S., Rodriguez, Y.: Torchvision the machine-vision package of torch. In: Pro- ceedings of the 18th ACM International Conference on Multimedia, pp. 1485–1488 (2010)

  14. [22]

    In: Indian Conference on Computer Vision, Graphics and Image Processing (2008)

    Nilsback, M.-E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian Conference on Computer Vision, Graphics and Image Processing (2008)

  15. [23]

    In: Advances in Neural Information Processing Systems (2019)

    Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. In: Advances in Neural Information Processing Systems (2019)

  16. [24]

    In: IEEE Conference on Computer Vision and Pattern Recognition (2012)

    Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.V.: Cats and dogs. In: IEEE Conference on Computer Vision and Pattern Recognition (2012)

  17. [25]

    In: International Conference on Machine Learning, pp

    Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, pp. 8748–8763 (2021). PmLR 18

  18. [26]

    arXiv preprint arXiv:2402.09668 (2024)

    Sachdeva, N., Coleman, B., Kang, W.-C., Ni, J., Hong, L., Chi, E.H., Caverlee, J., McAuley, J., Cheng, D.Z.: How to train data-efficient llms. arXiv preprint arXiv:2402.09668 (2024)

  19. [27]

    Settles, B.: Active learning literature survey (2009)

  20. [28]

    In: Proceedings of the 39th International Conference on Machine Learning (2022)

    Sorscher, B., Geiger, M., Bautista, X., Bordelon, B., Snyder, J., Sachs, S., Ganguli, S., Schapiro, A.C.: Beyond neural scaling laws: beating power law scaling via data pruning. In: Proceedings of the 39th International Conference on Machine Learning (2022)

  21. [29]

    International Journal of Computer Vision 130(6), 1526–1565 (2022)

    Soviany, P., Ionescu, R.T., Rota, P., Sebe, N.: Curriculum learning: A survey. International Journal of Computer Vision 130(6), 1526–1565 (2022)

  22. [30]

    arXiv preprint arXiv:1711.00489 (2017)

    Smith, S.L., Kindermans, P.-J., Ying, C., Le, Q.V.: Don’t decay the learning rate, increase the batch size. arXiv preprint arXiv:1711.00489 (2017)

  23. [31]

    arXiv preprint arXiv:1708.00489 (2017)

    Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017)

  24. [32]

    arXiv preprint arXiv:1801.01587 (2018)

    Shaham, U., Stanton, K., Li, H., Nadler, B., Basri, R., Kluger, Y.: Spectralnet: Spectral clustering using deep neural networks. arXiv preprint arXiv:1801.01587 (2018)

  25. [33]

    In: Proceedings of the IEEE International Conference on Computer Vision, pp

    Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness of data in deep learning era. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 843–852 (2017) Von Luxburg, U.: A tutorial on spectral clustering. Statistics and co...

  26. [34]

    IEEE transactions on pattern analysis and machine intelligence 44(9), 4555–4576 (2021)

    Wang, X., Chen, Y., Zhu, W.: A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence 44(9), 4555–4576 (2021)

  27. [35]

    In: Proceedings of the 24th International Conference on Machine Learning, pp

    Zhao, Z., Liu, H.: Spectral feature selection for supervised and unsupervised learning. In: Proceedings of the 24th International Conference on Machine Learning, pp. 1151–1157 (2007) 19

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.