Pith. sign in

REVIEW 5 major objections 5 minor 25 references

Robust Multimodal Learning via Entropy-Gated Contrastive Fusion

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 A single lightweight fusion layer, AECF, can keep multimodal models reliable and well calibrated when inputs go missing, without retraining frozen encoders.

desk verdict New combination, real gains over a naive gate, but the headline calibration guarantee is circular, one theorem is missing, and the ablation contradicts the core mechanism. read the letter →

arxiv 2505.15417 v1 pith:BLJI2TYT submitted 2025-05-21 stat.ML cs.LG

Leon Chlon , Maggie Chlon , MarcAntonio M. Awada This is my paper · ORCID
classification stat.MLcs.LG
keywords multimodalfusionmissingmodalityreliabilitycalibrationentropyregularizationcontrastivelearningcurriculummaskingmixtureofexpertsexpectederror
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 claims that a single, lightweight fusion layer called AECF can make multimodal models reliable when inputs go missing and keep their confidence scores calibrated, without retraining any pretrained encoders. The layer combines three mechanisms: a per-instance entropy penalty that adapts to model uncertainty, a contrastive loss that pushes confidence scores to respect the subset ordering (a subset of modalities should never be more confident than a superset), and a curriculum mask that learns to drop the modalities the gate relies on too heavily. The paper argues these mechanisms are complementary and proves two formal guarantees: a worst-case regret bound for the adaptive gate, and a PAC-style bound showing that the calibration loss keeps expected calibration error from growing as modalities are added. On AV-MNIST and MS-COCO, AECF reports masked-input accuracy gains of up to +18 percentage points at a 50% drop rate while roughly halving expected calibration error and adding under 1% runtime.

What carries the argument

The central object is the AECF fusion layer, a two-layer gate $g_\phi$ trained with the composite objective $\mathcal{L}=\mathcal{L}_{\text{task}}+\lambda(x)(-H(p))+\gamma\mathcal{L}_{\text{CEC}}+\beta\mathcal{L}_{\text{mask}}$. The entropy term $-H(p)$ with per-instance coefficient $\lambda(x)$ acts as an online log-barrier, and the convex conjugate of entropy over the simplex yields the $\log M/\lambda_{\min}$ regret bound. CEC is a squared-hinge loss $\text{ReLU}(c(A)-c(B))^2$ over subset pairs, a differentiable relaxation of isotonic regression on the subset lattice. ACM selects masks via a softmax over gate entropy $H(p(x\setminus S))/\eta$, an $O(M)$ approximation to the exponential $2^M$ teacher-student maximization. Together, these three components are what carry both the robustness and the calibration claims.

What would settle it

Train AECF on a three-modality dataset and, after training, measure confidences $c(A)$ and $c(B)$ for every pair of modality subsets $A\subset B$. If any inversion $c(A)>c(B)$ persists, or if worst-subset ECE increases when an extra modality is added, the monotone-calibration claim is falsified.

Watch

Extended reading notes

Core claim

AECF is claimed to be the first single fusion layer that is simultaneously instance-adaptive, lattice-calibrated, and curriculum-aware. Its meta-adaptive entropy gate computes a per-sample coefficient $\lambda(x)$ from MC-dropout or ensemble variance, tightening entropy regularization exactly on uncertain inputs. Contrastive expert calibration (CEC) penalizes squared-hinge violations of $c(A)\le c(B)$ for $A\subset B$ over all $2^M-1$ modality subsets, and adaptive curriculum masking samples missing-modality patterns in proportion to gate entropy, adversarially probing dominant modalities. The paper's formal results are a worst-case subset regret bound of $\log M/\lambda_{\min}$ plus $O(1/\sqrt{T})$ optimization terms, and a PAC bound for CEC that, under a no-inversion condition, implies expected calibration error cannot increase when modalities are added. Empirically, on MS-COCO AECF improves masked-input mAP by up to +18 pp at a 50% drop rate, reduces ECE by up to 2x relative to fixed-weight baselines, and adds less than 1% runtime with frozen backbones.

Load-bearing premise

The monotone-calibration guarantee assumes the model already ranks confidence correctly, $c(A)\le c(B)$ for $A\subset B$, almost surely, and the proof that CEC enforces this ranking assumes the ranking already exists.

Editorial extensions

If this is right

  • AECF can be inserted in front of any frozen multimodal encoder stack and improve masked-input reliability without retraining the encoders.
  • At 50% test-time modality dropout, AECF reports up to +18 pp mAP over fixed-weight fusion on MS-COCO while keeping ECE substantially lower.
  • If the no-inversion condition holds, adding a modality can never increase worst-subset expected calibration error, making calibration compositional across modality subsets.
  • The regret bound $\log M/\lambda_{\min}$ means per-input uncertainty estimates directly tighten the worst-case loss over missing subsets, concentrating regularization where the risk is highest.
  • The curriculum mask prevents gate entropy collapse; the ablation shows removing it costs 9-11 pp mAP under masking with no ECE benefit.

Reading between the lines

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

  • An extension the paper leaves implicit is that if CEC genuinely drives the model into the no-inversion region, Corollary 3 could be restated without its assumption; a training procedure that explicitly projects confidences onto the subset lattice would make monotone calibration unconditional.
  • The O(M) mask sampling and per-instance $\lambda(x)$ suggest AECF could scale to settings with many modalities, where enumerating all subsets is infeasible; this is testable on a high-modality dataset the paper does not include.
  • The regret analysis treats the curriculum term as a smooth perturbation; a complementary account could model the teacher-student mask game as a bandit problem and derive a regret bound for mask selection itself.
  • Because $\lambda(x)$ is computed from MC-dropout or ensemble variance, other uncertainty estimators could be substituted; whether the guarantees survive likely depends on the estimator's smoothness, which the paper does not analyze.
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 AECF, a multimodal fusion layer combining three modules: a per-instance adaptive entropy-regularized gate, a contrastive expert calibration (CEC) loss over modality subsets, and an entropy-driven curriculum mask. The authors claim that AECF simultaneously provides robustness and calibration under missing modalities, with formal guarantees of a worst-case subset regret bound and a PAC bound ensuring ECE cannot increase as modalities are added. Experiments on AV-MNIST and MS-COCO report improved masked-input mAP and reduced ECE with under 1% runtime overhead.

Significance. If the claims were substantiated, AECF would be a practically useful contribution: it keeps encoders frozen, is lightweight, and targets a realistic failure mode (missing modalities) while addressing calibration. The experimental design includes full-model ablations on two benchmarks, which is a strength. However, the central formal calibration guarantee is not proven, the adaptive entropy module is contradicted by the paper's own ablation table at high missing rates, and the headline ECE reduction is arithmetically impossible. As written, the evidence does not support the advertised contributions.

major comments (5)
  1. [§4.3, Corollary 3] Corollary 3 assumes the no-inversion condition c(A) ≤ c(B) almost surely whenever A⊂B, which is exactly the monotone ranking property that CEC is claimed to enforce. The paper provides no argument that the squared-hinge loss in Eq. (3) produces this condition: the loss is applied to sampled subset pairs and is a soft penalty with no margin or guarantee on all subsets. Moreover, Proposition 2, cited in Sections 1 and 3.3 as the proof that CEC decreases positive-ECE on the subset lattice, does not appear anywhere in the manuscript. Theorem 1 also does not close this gap, because it upper-bounds max_S ECE by a statistical term, an optimization term, and ECE(g⋆_S) of an unspecified lattice-isotonic optimum, without constructing calibrated scores that satisfy no-inversion. Thus the advertised guarantee that 'ECE cannot increase as modalities are added' is assumed rather than derived.
  2. [§5.4, Table 2] The ablation results contradict the claim that the entropy module is essential. At rnd50, the no-entropy ablation has higher mAP (0.443 vs. 0.440 for AECF) and lower ECE (0.018 vs. 0.020); at rnd30 it also has lower ECE (0.013 vs. 0.014). The text in §5.4 states that removing the entropy term 'increases ECE by nearly 2×', but the table shows the opposite. Since the adaptive entropy gate is one of the three headline contributions, this is a load-bearing inconsistency that must be addressed.
  3. [Abstract, §5.3] The abstract's claim of 'reducing ECE by up to 200%' is arithmetically impossible for ECE, a nonnegative metric; a reduction by more than 100% would require ECE to become negative. The actual numbers in Table 2 show at most a reduction from 0.028 to 0.020 at rnd50 (about 29%) and from 0.011 to 0.009 on full input. This headline claim should be corrected or removed.
  4. [Table 1] The AV-MNIST results table contains nonsensical entries. For the Image-only row, 'Acc full' is 99.97 while 'Acc img' is 78.6, although an image-only model's accuracy cannot differ between a full input and an image-only input. Similarly, Audio-only reports 100 for 'Acc aud' but has no meaningful 'Acc full' discrepancy; Equal fuse reports 'Acc aud' 100 and 'Acc img' 55.4, which is unexplained. The column definitions and the evaluation protocol need to be stated precisely, and the table needs to be corrected.
  5. [Theorem 3, Lemma 4] The worst-case subset regret bound max_S [R_T(S) − R_T(∅)] ≤ log M / λ_min + O(1/√T) is a restatement of the standard Hedge bound for entropy-regularized experts. The adaptive λ(x) enters only through the global lower bound λ_min, so no formal benefit of per-instance adaptation is shown. This contradicts the claim in §3.2 that the adaptive coefficient 'provably shrinks worst-subset regret'; the proof as written does not establish any dependence on the adaptive mechanism.
minor comments (5)
  1. [§2] The Related Work section has a broken sentence: 'Recent calibration unexplored and incurring significant compute.' is incomplete, and the paragraph preceding 'Positioning of AECF' ends abruptly with '[Wang et al., 2025] in that both keep encoder gradients...' missing a main clause.
  2. [§3.3] The text refers to 'Table??' in 'Empirically this reduces worst-subset ECE by 30–40% on all benchmarks (Table??)', leaving an unresolved placeholder.
  3. [Algorithm 2] Algorithm 2 is labeled 'Algorithm 2' but no Algorithm 1 appears; additionally, in line 4 the notation 'S∼Bernoulli(1−π_t)' is ambiguous because S is a subset, not a single Bernoulli variable, and the per-modality masking distribution should be specified.
  4. [Appendix D] Appendix D references 'Eq. (3.4)' for the closed-form curriculum mask, but no such equation number exists; the expression π_t(S) ∝ exp(H(p_t(x\S))/η) is not numbered.
  5. [§1] The text contains the typo 'Fig. Fig. 1' in the description of the AECF pipeline.

Circularity Check

2 steps flagged · score 7.0 of 10

Monotone-calibration guarantee is assumed, not derived: Corollary 3's no-inversion precondition is exactly the property CEC claims to enforce.

  1. self definitional [Section 4.3, Corollary 3]
    "Corollary 3 (Monotone calibration). Under the no-inversion condition c(A) ≤ c(B) almost surely whenever A⊂B, running CEC with the step size in Theorem 1 gives a non-increasing sequence of maxA ECE(A) and hence preserves calibration monotonicity."

    The no-inversion condition is precisely the ranking consistency that CEC is advertised to enforce: Section 4.1 states 'if A⊂B the model with more information should not appear less confident' and says CEC 'guarantees ECE cannot increase as additional modalities become available (§4.3)'. Corollary 3 therefore assumes the conclusion. The squared-hinge loss of Eq. (3) is a soft penalty over sampled subset pairs; the paper gives no argument that minimizing it produces the almost-sure no-inversion condition, or that it implies ECE monotonicity across all subsets. Without that argument, the central calibration guarantee is an assumption, not a derived result.

  2. other [Section 1, Contribution 2; Section 3.3 'Guarantee']
    "Proposition 2 proves CEC decreases positive-ECE on the entire subset lattice. ... Under mild assumptions each update decreases the maximum expected calibration error across subsets (proof sketch in §4.3)."

    The only stated proof of CEC's calibration effect is Proposition 2, which never appears anywhere in the manuscript. The referenced proof sketch in §4.3 is Corollary 3, which assumes the no-inversion condition as its precondition. Thus the claimed calibration guarantee is not independently derived; it reduces to an unstated assumption and an absent proposition, leaving the advertised 'PAC bound for CEC that ensures ECE cannot increase as modalities are added' unsupported.

full rationale

The paper's central calibration claim is circular. Corollary 3 states that, under the no-inversion condition c(A) ≤ c(B) almost surely whenever A⊂B, running CEC gives a non-increasing sequence of max_A ECE(A). That condition is exactly the monotone confidence ranking that CEC is claimed to enforce via the pairwise squared-hinge loss in Eq. (3). No proof shows that minimizing Eq. (3) creates the almost-sure no-inversion property; the soft penalty samples finitely many pairs and has no margin guarantee. The paper also cites Proposition 2 for the claim that CEC decreases positive-ECE on the subset lattice, but no Proposition 2 exists in the manuscript. Theorem 1's PAC bound does not repair the chain: it upper-bounds max_S ECE by a statistical term, an optimization term, and ECE(g⋆_S) of an unspecified lattice-isotonic optimum, without constructing calibrated scores that satisfy no-inversion or showing that ECE is monotone across subsets. By contrast, the worst-case regret bound in Theorem 3 is the standard Hedge bound log M / λ_min, as the paper itself acknowledges via Shalev-Shwartz and Cesa-Bianchi-Lugosi; it is a known result rather than a circular step, though the adaptive λ(x) does not appear in the bound. The empirical mAP and ECE comparisons on AV-MNIST and MS-COCO are independent of the circular calibration proof and support the method's practical robustness, so the paper is only partially circular. Score 7 reflects that one of the two headline formal guarantees, monotone calibration, reduces to an assumption rather than a derivation.

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

The central claim rests on validation-tuned hyperparameters (λmax, πmax, vmax), convexity and Lipschitz assumptions that do not hold for deep classifiers, and a circular no-inversion condition in Corollary 3. No new physical entities are introduced.

free parameters (5)
  • λmax (maximum entropy coefficient) = 0.08
    Tuned on the COCO validation split; reported best compromise for robustness and full-input accuracy, so the final results depend on this validation-selected value.
  • πmax (maximum curriculum dropout rate) = 0.40
    Tuned on the COCO validation split among {0.4, 0.5}; the reported results use 0.40.
  • λmin (minimum entropy coefficient) = unspecified
    The standing assumption requires λmin > 0, and Eq. (2) sets λ(x) = λmin + softplus(...), but its numerical value is never given. The convexity and regret bounds depend on it.
  • vmax (validation-set maximum variance) = measured on validation set
    Used to clip the softplus argument and define λmax = λmin + softplus(vmax); it is a data-dependent constant.
  • η (ACM temperature) = unspecified
    The closed-form mask distribution πt(S) ∝ exp(H(pt)/η) depends on a temperature η that is not specified in the paper.
assumptions (6)
  • domain assumption Task loss ℓ is convex, 1-Lipschitz, and bounded in [0,1]
    Stated in Appendix A and used in Lemma 1 and Theorem 3. Real cross-entropy losses for deep classifiers are not 1-Lipschitz in the network parameters and are not convex in general.
  • domain assumption Head hψ is linear in its input
    Appendix A. Needed for convexity of the fused predictor. The paper later says a linear classifier is trained, but the head operates on the fused representation, and the claim of linearity in p is not established.
  • domain assumption Each encoder fm is σ-Lipschitz
    Appendix A. Used to bound the predictor Lipschitz constant; plausible but unverified for the frozen CLIP and CNN encoders.
  • domain assumption Per-batch loss ft is L-Lipschitz and β-smooth, and the feasible set Θ has diameter D
    Theorem 2. Standard optimization assumptions, but not verified for the actual non-convex neural network objective.
  • ad hoc to paper No-inversion condition c(A) ≤ c(B) almost surely whenever A⊂B
    Corollary 3 assumes the very monotone calibration property CEC is supposed to enforce, making the calibration guarantee circular.
  • standard math Slater's condition and strong duality for the entropy-regularized simplex objective
    Lemma 1 and Appendix B.1. Standard for strictly convex objectives on the simplex with non-empty relative interior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Multimodal Learning via Entropy-Gated Contrastive Fusion." pith.science (2026). https://pith.science/paper/BLJI2TYT

@misc{pith2026250515417,
  author       = {Pith},
  title        = {Pith review of: Robust Multimodal Learning via Entropy-Gated Contrastive Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BLJI2TYT}},
  note         = {Machine review of arXiv:2505.15417}
}
read the original abstract

Real-world multimodal systems routinely face missing-input scenarios, and in reality, robots lose audio in a factory or a clinical record omits lab tests at inference time. Standard fusion layers either preserve robustness or calibration but never both. We introduce Adaptive Entropy-Gated Contrastive Fusion (AECF), a single light-weight layer that (i) adapts its entropy coefficient per instance, (ii) enforces monotone calibration across all modality subsets, and (iii) drives a curriculum mask directly from training-time entropy. On AV-MNIST and MS-COCO, AECF improves masked-input mAP by +18 pp at a 50% drop rate while reducing ECE by up to 200%, yet adds 1% run-time. All back-bones remain frozen, making AECF an easy drop-in layer for robust, calibrated multimodal inference.

Figures

Figures reproduced from arXiv: 2505.15417 by the authors.

Figure 1
Figure 1. AECF pipeline. A two–layer gate mixes frozen encoder features. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Per-sample gate entropy H(p) versus model confidence (maxk σ(logitk )). Only the full AECF model (a) shows the expected monotone relationship, corroborating the theory of §4.3. Gate ablation. Equal averaging fails under missing inputs and is out￾performed by all adaptive variants—even single-modality baselines beat it at π = 0.5. 5.5 Gate behaviour and cost [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages

  1. [1]

    Alfasly, E

    M. Alfasly, E. Erzin, and G. Varol. Learnable modality dropout for audio–visual classification. In European Conference on Computer Vision (ECCV) Workshops, 2022

    work page 2022

  2. [2]

    Aydore, L

    S. Aydore, L. Dicker, and D. Foster. A local regret in nonconvex online learning, 2018

    work page 2018

  3. [3]

    Z. Cao, X. Li, and L. Zhu. Infrared–visible image fusion via multi–modal mixture of experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4567--4576, 2023

    work page 2023

  4. [4]

    Cesa-Bianchi and G

    N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press, Cambridge, UK, 2006. ISBN 9780521841085. doi:10.1017/CBO9780511546921

    doi:10.1017/cbo9780511546921 2006

  5. [5]

    Fedus, B

    W. Fedus, B. Zoph, and N. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961, 2021

    arXiv 2021

  6. [6]

    Gal and Z

    Y. Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning (ICML), pages 1050--1059, 2016

    work page 2016

  7. [7]

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), pages 1321--1330, 2017

    2017

  8. [8]

    Hallak, S

    A. Hallak, S. Ghiassian, M. Dimakopoulou, and S. Mannor. Regret minimization in stochastic non-convex learning via a projected measure. In International Conference on Machine Learning, pages 4014--4024, 2021

    work page 2021

Show all 25 references
  1. [9]

    Y. Han, J. Chen, and A. Zisserman. Fusemoe: Flexible mixture-of-experts for multimodal transformers. In Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  2. [10]

    Hazan, K

    E. Hazan, K. Singh, and C. Zhang. Efficient regret minimization in non-convex games. In International Conference on Machine Learning, pages 1427--1436, 2017

    2017

  3. [11]

    M. I. Jordan and R. A. Jacobs. Hierarchical mixtures of experts and the EM algorithm. Neural Computation, 6 0 (2): 0 181--214, 1994

    1994

  4. [12]

    A. Kuhn, A. Varamesh, E. H \"u llermeier, and M. Eickenberg. Fusemoe: Learning to fuse modalities with conditional mixture of experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. URL https://arxiv.org/abs/2306.05466. arXiv:2306.05466

    2023 arXiv

  5. [13]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, and C. Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems (NeurIPS), pages 6405--6416, 2017

    2017

  6. [14]

    M. Ma, Q. Huang, and D. Tao. Calibrating multimodal learning. In Proceedings of the 40th International Conference on Machine Learning (ICML), 2023

    2023

  7. [15]

    Moulines and F

    E. Moulines and F. Bach. Non-asymptotic analysis of stochastic approximation algorithms for machine learning. In Advances in Neural Information Processing Systems (NeurIPS), volume 24, pages 451--459, 2011

    2011

  8. [16]

    Mukhoti, Y

    J. Mukhoti, Y. Gal, and P. H. S. Torr. Dense focal loss improves calibration for dense prediction. In Proceedings of the 41st International Conference on Machine Learning (ICML), 2024

    2024

  9. [17]

    Neverova, C

    N. Neverova, C. Wolf, G. Taylor, and F. Nebout. ModDrop : Adaptive multi-modal gesture recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 756--763, 2015

    2015

  10. [18]

    Nezakati and W

    M. Nezakati and W. Xie. Masked modality projection for robust multimodal learning. arXiv preprint arXiv:2403.01234, 2024

    2024 arXiv

  11. [19]

    Ramachandram and G

    D. Ramachandram and G. W. Taylor. Deep multimodal representation learning: A survey. IEEE Signal Processing Magazine, 34 0 (6): 0 96--108, 2017. doi:10.1109/MSP.2017.2738401. URL https://ui.adsabs.harvard.edu/abs/2017ISPM...34...96R/abstract

    2017

  12. [20]

    Reza and R

    I. Reza and R. Timofte. Uni–adapter: Parameter–efficient missing–modality adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345--12355, 2023

    2023

  13. [21]

    S. J. N. Robert A. Jacobs, Michael I. Jordan and G. E. Hinton. Adaptive mixtures of local experts. Neural Computation, 3 0 (1): 0 79--87, 1991

    1991

  14. [22]

    Shalev-Shwartz

    S. Shalev-Shwartz. Online learning and online convex optimization. Foundations and Trends in Machine Learning, 4 0 (2): 0 107--194, 2012. doi:10.1561/2200000018

    2012 doi

  15. [23]

    Shazeer, A

    N. Shazeer, A. Mirhoseini, K. Maziarz, et al. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  16. [24]

    Y. Tang, E. Rohaninejad, and J. Yu. Relative calibration for multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  17. [25]

    Y. Wang, S. Venkataramani, and T. Darrell. Hyperformer: Hypernetworks for robust missing-modality fusion. In Proc. NeurIPS, 2025

    2025

Pith tools

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