Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Consistency-aware Fake Videos Detection on Short Video Platforms

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A new model detects fake news videos by scoring how consistently their visual, textual, and audio tracks agree, reporting state-of-the-art accuracy on two benchmarks.

desk verdict A clever consistency-modelling idea is undermined by pseudo-labels that are partly set from the ground truth, so the reported gains may be label leakage. read the letter →

arxiv 2504.21495 v1 pith:GB3DV6ZH submitted 2025-04-30 cs.CV cs.MM

Junxi Wang , Jize liu , Na Zhang , Yaxiong Wang This is my paper · ORCID
classification cs.CVcs.MM
keywords fakenewsdetectionshortvideoplatformscross-modalconsistencymultimodallargelanguagemodelpseudo-labelgenerationfeaturefusionSVTT
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 argues that manipulated news videos often contain mismatches between what is shown, what is written, and what is narrated, and that existing detectors ignore these contradictions. It proposes CA-FVD, which uses a multimodal large language model to generate pseudo-labels for pairwise modality agreement, trains a cosine-similarity loss that penalizes inconsistent pairs, and then fuses per-modality probability scores for the final verdict. On the FakeSV and FakeTT benchmarks, the model reports 85.79% and 81.61% accuracy, beating the previous best FakingRecipe method by 1.10 and 2.35 points. The central claim is that explicitly modeling cross-modal consistency, rather than only fusing features, is what drives the gain.

What carries the argument

The central mechanism is the cross-modal consistency learning loop: an MLLM (MiniCPM-o) turns open-ended questions about whether the visual, textual, and audio information of a video align into binary pseudo-labels, and a cosine-similarity loss on [CLS] tokens of the three modalities, weighted by those pseudo-labels, forces the encoders to make their representations agree for real videos and disagree for fakes. This converts an unannotated property (modality mismatch) into a supervised objective, which the paper claims is what lets the model exploit contradictions that existing fusion-only detectors miss. The second half of the machinery is probability-score fusion, which multiplies the pairwise and emotion branch probabilities rather than simply concatenating features.

What would settle it

Retrain CA-FVD on the same FakeSV and FakeTT splits but generate the consistency pseudo-labels purely from the MLLM prompts, dropping the ground-truth adjustment described in Section 4.1; if the accuracy on either dataset falls back to the level of the no-consistency ablation (85.06 on FakeSV, 78.93 on FakeTT), the reported gains are attributable to label leakage rather than to consistency modeling.

Watch

Extended reading notes

Core claim

The paper's central discovery is that cross-modal inconsistency is a usable discriminative signal for fake news video detection. The authors design a two-stage pipeline: in Cross-modal Consistency Learning, prompts are fed with the video and text into MiniCPM-o to obtain 0/1 pseudo-labels for each of the three modality pairs (vision-text, audio-text, vision-audio); the model's [CLS] tokens for the three modalities are then compared with a cosine loss that pulls consistent pairs together and pushes inconsistent pairs apart. In the Multi-modal Collaborative Diagnosis stage, co-attention and a Transformer fuse the semantic streams, and the probability scores from the three pairwise branches are multiplied together with an emotion branch to produce the final verdict. The paper reports state-of-the-art accuracy on both datasets, and its ablation studies show that removing either the consistency module or the score-fusion strategy reduces accuracy on both benchmarks, with larger drops on FakeTT.

Load-bearing premise

The load-bearing premise is that the MLLM's consistency pseudo-labels can be corrected with the ground-truth fake/real label without leaking the answer into the consistency supervision, so that the reported accuracy gains reflect genuine cross-modal inconsistency rather than memorization of the target label.

Editorial extensions

If this is right

  • If cross-modal consistency is a reliable cue, fake-video creators must either hide mismatches across all three modalities or risk detection, raising the cost of producing convincing fakes.
  • The MLLM-based pseudo-labeling removes the need for manually annotated modality-matching labels, making the method applicable to datasets that only have binary fake/real labels.
  • The ablations indicate that probability-score fusion and feature fusion are complementary; combining verdict-level and feature-level evidence appears to be a stronger design than either alone.
  • The reported performance on both a Chinese-language and an English-language benchmark suggests the consistency cue is not language-specific.

Reading between the lines

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

  • Inference: The same consistency-diagnosis recipe could be transferred to other modality triples, such as image-text-social-context rumor detection, where content mismatches are common and manually annotated alignment labels are unavailable.
  • Inference: If the MLLM's consistency judgments are trustworthy, they could become a pretraining objective for video-language models, improving general video understanding rather than only fake-news detection.
  • Inference: A learned gating mechanism over the pairwise probability scores could replace the fixed multiplication in Eq. (6), letting the model down-weight branches whose modalities are uninformative for a given video.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CA-FVD, a fake-news-video detector for short-video platforms. The method has two main modules: Cross-modal Consistency Learning (CMCL), which uses an MLLM to generate pseudo-labels for pairwise visual-textual-audio consistency and then penalizes cosine-similarity mismatches between [CLS] tokens of different modalities, and Multi-modal Collaborative Diagnosis (MMCD), which fuses semantic and emotional features through co-attention and a Transformer and combines sub-scores from different branches. Experiments on FakeSV and FakeTT report Accuracy of 85.79 and 81.61, respectively, outperforming the FakingRecipe baseline, together with ablations, hyperparameter sensitivity analysis, and case studies. The authors also release code.

Significance. If the cross-modal consistency signal were genuine, the idea of using an MLLM's open-world knowledge to supervise consistency between vision, text, and audio would be a useful contribution, and the public code release is a plus for reproducibility. However, the experimental support for the central claim is undermined by the pseudo-label adjustment procedure in Section 4.1, which injects the ground-truth authenticity label into the consistency supervision. As a result, the reported accuracy gains cannot be attributed to consistency modeling rather than target-label leakage. The lack of error bars and the absence of an ablation that removes the ground-truth adjustment further weaken the claims.

major comments (4)
  1. [Section 4.1, Eqs. (3)-(4)] The pseudo-labels supervising the consistency loss are adjusted using the ground-truth authenticity label. The manuscript states, 'for genuine news videos, we default all three modality matching labels to true.' Therefore, for every real video the three consistency targets are set to 'consistent' regardless of the actual inter-modal semantic agreement, and the consistency loss in Eq. (4) can be minimized by aligning modality representations whenever the ground-truth label is 'real'. This makes the CMCL branch behave as an additional authenticity classifier rather than a detector of inter-modal inconsistency. Because the reported gains over FakingRecipe are modest (85.79 vs. 84.69 on FakeSV; 81.61 vs. 79.26 on FakeTT), the target-conditioned supervision could explain the improvement without any genuine consistency modeling. The paper does not report the raw MLLM pseudo-label accuracy, nor does it ablate the ground-truth adjustment separately, so the central claim is not supported as presented.
  2. [Section 4.2, Table 2] All results are single-run point estimates without standard deviations, number of random seeds, or significance tests. The headline improvements over the strongest baseline are 1.10 accuracy points on FakeSV and 2.35 accuracy points on FakeTT, which are small enough that seed variability could change the ranking. The authors should report mean and standard deviation over multiple runs and perform significance tests for the main comparisons and for the ablations in Table 3.
  3. [Section 4.3, Table 3] The ablation study removes components such as the CLS token, individual consistency losses, emotional features, and score fusion, but it never removes the ground-truth adjustment of the pseudo-labels. Consequently, the contribution of the consistency branch is confounded with direct authenticity supervision. An ablation that uses the raw MLLM pseudo-labels without ground-truth adjustment is necessary to determine whether CMCL learns inter-modal consistency at all; without this condition, the accuracy gains attributed to the feature alignment module in Section 4.3 are not interpretable.
  4. [Section 3.3, Eqs. (6)-(7)] The final prediction is defined as a product of sub-scores (P = p_v * p_a * p_t * p_e) followed by P = P_total * softmax(P_total). If the sub-scores are probabilities, their product is not a calibrated probability and the additional softmax factor is dimensionally and semantically unusual; if they are raw logits, the product has no probabilistic interpretation. The output activation of each branch and the justification for this fusion formula should be specified, since the final detection decision depends on it.
minor comments (5)
  1. [Table 2] There are typos in the table headers: 'Accurancy' should be 'Accuracy' and 'Marco F1' should be 'Macro F1'; additionally, 'Adma optimizer' in Section 4.1 should be 'Adam optimizer'.
  2. [Throughout] Many equations contain missing placeholder symbols (e.g., '?????') that obscure the definitions of features, losses, and probability scores; the manuscript should be carefully proofread so that every symbol is rendered and defined on first use.
  3. [Section 3.2 / 3.3] The module name is given as 'Multimodal Feature Fusion (MFF)' in the abstract and elsewhere as 'MMF'; please use one consistent abbreviation throughout.
  4. [Section 4.1] The pseudo-label generation step records consistent labels as 0 and inconsistent labels as 1, but the ground-truth adjustment is described as defaulting labels to 'true'; please clarify the mapping between the boolean description and the numeric coding used in Eq. (3).
  5. [Reference [29]] The MiniCPM-o reference is a GitHub URL with no author, year, or version details; please cite a stable archival version or provide complete bibliographic information.

Circularity Check

1 steps flagged · score 7.0 of 10

Consistency pseudo-labels are constructed from the authenticity label, so the claimed consistency signal is not independent of the target.

  1. self definitional [Section 4.1 (Experimental Setup), used by Eq. (3), Eq. (4), and the overall loss Eq. (11)]
    "To ensure the accuracy of the pseudo-labels, we adjust them based on the ground truth labels. Specifically, for genuine news videos, we default all three modality matching labels to true."

    The pseudo-labels are the supervision for the cross-modal consistency loss: Eq. (3) uses the pseudo-label p_vt to choose the cosine-loss form, Eq. (4) sums the three pairwise consistency losses, and Eq. (11) adds L_cmcl to the total training objective. For every genuine video, the paper defines all pairwise matching labels as true by the ground-truth authenticity label, rather than by measured modality agreement. Therefore, the CMCL branch is trained, at least for the real class, to reproduce the target label through the auxiliary consistency loss; minimizing L_cmcl can be achieved by labeling genuine videos as consistent and fake videos as inconsistent, independent of actual inter-modal semantic agreement.

full rationale

The central conceptual claim is that explicitly modeling cross-modal consistency improves fake-video detection. That claim rests on the CMCL module, whose supervision is the pseudo-labels from Section 4.1. The quoted implementation sentence shows that, for genuine news videos, the three pairwise consistency labels are set to true by the ground-truth authenticity label. Since Eq. (3)-(4) feed these labels into L_cmcl and Eq. (11) adds L_cmcl to the total loss, the target label enters the training objective through the 'consistency' pathway by construction. This is a self-definitional step in the experimental support: the consistency signal is partly defined in terms of the very label the model is supposed to predict. The rest of the architecture—co-attention fusion, emotion features, probability score fusion—is evaluated against external baselines and is not circular, but the central consistency mechanism is not independently validated. A score of 7 reflects one load-bearing circular step that contaminates the main empirical claim, while acknowledging that the paper is not a complete derivation collapse and includes substantial non-circular components.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unvalidated assumptions. The most damaging is the ground-truth adjustment of pseudo-labels, which conflates the consistency signal with the target label. The hyperparameters are fitted on the test set without a clear validation split, and the score fusion heuristic is not derived from probability theory.

free parameters (3)
  • alpha = 0.5
    Loss-balancing hyperparameter selected via sensitivity analysis (Section 4.4), without a held-out validation protocol.
  • beta = 0.1
    Loss-balancing hyperparameter selected via sensitivity analysis (Section 4.4), without a held-out validation protocol.
  • gamma = 3.0
    Loss-balancing hyperparameter selected via sensitivity analysis (Section 4.4), without a held-out validation protocol.
assumptions (4)
  • domain assumption Cross-modal inconsistency is a reliable discriminative cue for fake news videos
    The entire method is built on the premise that mismatched modalities signal manipulation, stated in the Introduction and Figure 1.
  • ad hoc to paper MLLM-generated pseudo-labels, after ground-truth adjustment, accurately reflect multimodal consistency
    Section 4.1 adjusts pseudo-labels using ground truth labels, assuming this yields accurate consistency labels while keeping the evaluation valid.
  • domain assumption Pre-extracted features from FakingRecipe are valid representations for vision, text, and audio
    Section 4.1 states the model uses 'pre-extracted relevant features from existing work [2],' relying on those features without independent validation.
  • ad hoc to paper The product combination of probability scores in Eq. 6 produces a valid detection score
    Equation 6 multiplies raw probability scores without normalization or probabilistic derivation, yet it is used to produce the final prediction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Consistency-aware Fake Videos Detection on Short Video Platforms." pith.science (2026). https://pith.science/paper/GB3DV6ZH

@misc{pith2026250421495,
  author       = {Pith},
  title        = {Pith review of: Consistency-aware Fake Videos Detection on Short Video Platforms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GB3DV6ZH}},
  note         = {Machine review of arXiv:2504.21495}
}
read the original abstract

This paper focuses to detect the fake news on the short video platforms. While significant research efforts have been devoted to this task with notable progress in recent years, current detection accuracy remains suboptimal due to the rapid evolution of content manipulation and generation technologies. Existing approaches typically employ a cross-modal fusion strategy that directly combines raw video data with metadata inputs before applying a classification layer. However, our empirical observations reveal a critical oversight: manipulated content frequently exhibits inter-modal inconsistencies that could serve as valuable discriminative features, yet remain underutilized in contemporary detection frameworks. Motivated by this insight, we propose a novel detection paradigm that explicitly identifies and leverages cross-modal contradictions as discriminative cues. Our approach consists of two core modules: Cross-modal Consistency Learning (CMCL) and Multi-modal Collaborative Diagnosis (MMCD). CMCL includes Pseudo-label Generation (PLG) and Cross-modal Consistency Diagnosis (CMCD). In PLG, a Multimodal Large Language Model is used to generate pseudo-labels for evaluating cross-modal semantic consistency. Then, CMCD extracts [CLS] tokens and computes cosine loss to quantify cross-modal inconsistencies. MMCD further integrates multimodal features through Multimodal Feature Fusion (MFF) and Probability Scores Fusion (PSF). MFF employs a co-attention mechanism to enhance semantic interactions across different modalities, while a Transformer is utilized for comprehensive feature fusion. Meanwhile, PSF further integrates the fake news probability scores obtained in the previous step. Extensive experiments on established benchmarks (FakeSV and FakeTT) demonstrate our model exhibits outstanding performance in Fake videos detection.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Detecting AI-Generated Video: A Vision-Language Dual-View Survey

    cs.CV 2026-07 conditional novelty 6.0 of 10

    AIGC-V detection should be treated as factual fidelity verification and organized by a four-layer vision-language dual-view taxonomy spanning cues, motion, cross-modal consistency, and world-level reasoning.

  2. FakeSV-VLM: Taming VLM for Detecting Fake Short-Video News via Progressive Mixture-Of-Experts Adapter

    cs.MM 2025-08 reject novelty 5.0 of 10

    FakeSV-VLM reaches 90.22% and 89.30% accuracy on FakeSV and FakeTT by adding a two-stage MoE adapter and contrastive alignment to InternVL2.5-8B.

Reference graph

Works this paper leans on

29 extracted references · 20 canonical work pages · cited by 2 Pith papers

  1. [1]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Qi, P., Bu, Y., Cao, J., et al.: FakeSV: A multimodal benchmark with rich social context for fake news detection on short video platforms. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37(12), pp. 14444–14452 (2023)

    work page 2023

  2. [2]

    In: Proceedings of the 32nd ACM International Conference on Multimedia, pp

    Bu, Y., Sheng, Q., Cao, J., et al.: FakingRecipe: Detecting fake news on short video platforms from the perspective of creative process. In: Proceedings of the 32nd ACM International Conference on Multimedia, pp. 1351–1360 (2024)

    work page 2024

  3. [3]

    In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, vol

    Devlin, J., Chang, M.W., Lee, K., et al.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, vol. 1, pp. 4171–4186 (2019)

    work page 2019

  4. [4]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al.: An image is worth 16×16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    arXiv 2020

  5. [5]

    In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pp

    Choi, H., Ko, Y.: Using topic modeling and adversarial neural networks for fake news video detection. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pp. 2950–2954 (2021)

    work page 2021

  6. [6]

    In: 2021 IEEE International Conference on Big Data (Big Data), pp

    Shang, L., Kou, Z., Zhang, Y., et al.: A multimodal misinformation detector for COVID-19 short videos on TikTok. In: 2021 IEEE International Conference on Big Data (Big Data), pp. 899–908. IEEE (2021)

    work page 2021

  7. [7]

    arXiv preprint arXiv:2412.05271 (2024)

    Chen, Z., Wang, W., Cao, Y., et al.: Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271 (2024)

    arXiv 2024

  8. [8]

    arXiv preprint arXiv:2306.05241 (2023)

    Qi, P., Zhao, Y., Shen, Y., et al.: Two heads are better than one: Improving fake news video detection by correlating with neighbors. arXiv preprint arXiv:2306.05241 (2023)

    arXiv 2023

Show all 29 references
  1. [9]

    Li, X., Xiao, X., Li, J., et al.: A CNN-based misleading video detection model. Sci. Rep. 12(1) (2022)

    2022

  2. [10]

    In: ACL Workshop on Natural Language Processing for COVID-19 (NLP-COVID) (2020)

    Medina Serrano, J.C., Papakyriakopoulos, O., Hegelich, S.: NLP-based feature extraction for the detection of COVID-19 misinformation videos on YouTube. In: ACL Workshop on Natural Language Processing for COVID-19 (NLP-COVID) (2020)

    2020

  3. [11]

    In: Proceedings of the 32nd ACM International Conference on Multimedia, pp

    Liu, X., Li, P., Huang, H., et al.: FKA-OWL: Advancing multimodal fake news detection through knowledge-augmented LVLMs. In: Proceedings of the 32nd ACM International Conference on Multimedia, pp. 10154–10163 (2024)

    2024

  4. [12]

    arXiv preprint arXiv:2101.09810 (2021)

    Ghanem, B., Ponzetto, S.P., Rosso, P., et al.: FakeFlow: Fake news detection by modeling the flow of affective information. arXiv preprint arXiv:2101.09810 (2021)

    2021 arXiv

  5. [13]

    In: Disinformation, Misinformation, and Fake News in Social Media: Emerging Research Challenges and Opportunities, pp

    Cao, J., Qi, P., Sheng, Q., et al.: Exploring the role of visual content in fake news detection. In: Disinformation, Misinformation, and Fake News in Social Media: Emerging Research Challenges and Opportunities, pp. 141–161 (2020)

    2020

  6. [14]

    Comput.-Mediat

    Sundar, S.S., Molina, M.D., Cho, E.: Seeing is believing: Is video modality more powerful in spreading fake news via online messaging apps? J. Comput.-Mediat. Commun. 26(6), 301–319 (2021)

    2021

  7. [15]

    arXiv preprint arXiv:2409.12191 (2024)

    Wang, P., Bai, S., Tan, S., et al.: Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

    2024 arXiv

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Qi, P., Yan, Z., Hsu, W., et al.: Sniffer: Multimodal large language model for explainable out-of-context misinformation detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13052–13062 (2024)

    2024

  9. [17]

    arXiv preprint arXiv:2305.16355 (2023)

    Su, Y., Lan, T., Li, H., et al.: PandaGPT: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355 (2023)

    2023 arXiv

  10. [18]

    arXiv preprint arXiv:2402.12226 (2024)

    Zhan, J., Dai, J., Ye, J., et al.: AnyGPT: Unified multimodal LLM with discrete sequence modeling. arXiv preprint arXiv:2402.12226 (2024)

    2024 arXiv

  11. [19]

    In: The Web Conference 2025 (2025)

    Hong, R., Lang, J., Xu, J., et al.: Following clues, approaching the truth: Explainable micro-video rumor detection via chain-of-thought reasoning. In: The Web Conference 2025 (2025)

    2025

  12. [20]

    Discourse Commun

    Igwebuike, E.E., Chimuanya, L.: Legitimating falsehood in social media: A discourse analysis of political fake news. Discourse Commun. 15(1), 42–58 (2021)

    2021

  13. [21]

    Fong, B.: Analysing the behavioural finance impact of 'fake news' phenomena on financial markets: A representative agent model and empirical validation. Financ. Innov. 7(1), 53 (2021)

    2021

  14. [22]

    Bezbaruah, S., Dhir, A., Talwar, S., et al.: Believing and acting on fake news related to natural food: The influential role of brand trust and system trust. Br. Food J. 124(9), 2937–2962 (2022)

    2022

  15. [23]

    arXiv preprint arXiv:2310.19193 (2023)

    Niu, S., Shrestha, D., Ghimire, A., et al.: A survey on watching social issue videos among YouTube and TikTok users. arXiv preprint arXiv:2310.19193 (2023)

    2023 arXiv

  16. [24]

    In: Proceedings of The Web Conference 2021, pp

    Zhang, X., Cao, J., Li, X., et al.: Mining dual emotion for fake news detection. In: Proceedings of The Web Conference 2021, pp. 3465–3476 (2021)

    2021

  17. [25]

    In: International Conference on Machine Learning, PMLR, pp

    Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, PMLR, pp. 8748–8763 (2021)

    2021

  18. [26]

    arXiv preprint arXiv:1911.02116 (2019)

    Conneau, A., Khandelwal, K., Goyal, N., et al.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019)

    2019 arXiv

  19. [27]

    IEEE/ACM Trans

    Hsu, W.N., Bolte, B., Tsai, Y.H.H., et al.: HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Trans. Audio Speech Lang. Process. 29, 3451–3460 (2021)

    2021

  20. [28]

    Lu, J., Batra, D., Parikh, D., Lee, S.: ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Adv. Neural Inf. Process. Syst. 32 (2019)

    2019

  21. [29]

    https://github.com/OpenBMB/MiniCPM-o

    A GPT-4o level MLLM for vision, speech and multimodal live streaming on your phone. https://github.com/OpenBMB/MiniCPM-o

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.