Pith. sign in

REVIEW 2 major objections 4 minor 24 references

How Should Vision-Language-Action Models Use Proprioceptive State?

T0 review · 2 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper's controlled experiments yield a routing rule: a single robot-state frame should enter the vision-language side, while a short multi-frame history should be routed to the action head.

desk verdict A genuinely controlled comparison of state wiring in VLAs with a useful repeat-current control; the headline routing rule is plausible but leans on single-seed point estimates, and the paper deserves peer review. read the letter →

arxiv 2608.03052 v1 pith:6IOGKYLO submitted 2026-08-04 cs.RO

Yiren Zhao , Ziyang Chen , Ziyang Rao , Pengteng Li , He Zhang , Weiyu Guo , Yandong Guo , Rushi Dai This is my paper · ORCID
classification cs.RO
keywords vision-language-actionmodelsproprioceptivestateconditioninghistoryinjectioninterfaceflow-matchingpolicyclosed-loopmanipulationdesignguidelines
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

Vision-language-action (VLA) models almost all take in the robot's own body state — end-effector pose, base pose, gripper aperture — but they disagree on how to represent it, where to inject it, and how much history to give it. This paper tries to settle those choices with everything else held fixed: one flow-matching VLA backbone, one dataset, one action representation, one evaluation protocol, and five state interfaces built on top. Across 45 atomic kitchen-manipulation tasks and 20 longer composite tasks, the authors find that current state helps selectively, that a short eight-frame history is useful while long raw histories can hurt, and that the best injection site flips with temporal budget: a single frame works best through the vision-language side, while a short history works best when it directly conditions action generation. The result is offered as a testable design rule, with the authors noting that most comparisons rest on a single training seed and that the rule is a directional pattern rather than an established causal mechanism.

What carries the argument

The load-bearing object is a set of five matched state interfaces implemented in one fixed flow-matching VLA scaffold: a discrete state prompt (state quantized into text tokens), a VLM prefix (continuous state tokens inserted into the vision-language context), an action prefix (state tokens placed in the action expert's causal suffix), a state expert (a dedicated state-processing stream), and feature modulation (per-feature scale/shift conditioning through cross-attention). They share the same backbone, data, action space, and training recipe, so differences in closed-loop success rate are attributed to the state design itself. The control that carries the temporal claim is slot-matched: an eight-frame history is compared against eight copies of the current state with image, instruction, slot count, and flow noise fixed, separating temporal content from added capacity. Paired task-bootstrap intervals provide the only interval-supported statements, and fixed-checkpoint probes of attention redistribution and flow-trajectory corrections trace the data-flow difference between VLM-side and action-side routes.

What would settle it

Re-run the central comparisons (one-frame and eight-frame action prefix versus VLM prefix, plus the slot-matched repeat-current control) on the same 20 composite tasks with several independent training seeds per interface; if the single-frame VLM-side advantage and the eight-frame action-side advantage do not reproduce across seeds, or if repeated current state matches genuine history once seed noise is accounted for, the routing rule fails.

Watch

Extended reading notes

Core claim

The central claim is that proprioceptive state should be routed according to how much history is available: with one frame, state belongs in the vision-language backbone (as a discrete state prompt or a continuous VLM prefix), because it can redistribute multimodal attention before action generation; with a short ordered history, state belongs at the action-generation module (as an action prefix), because that is where ordered temporal variation is exploited most reliably. On the 20 composite tasks the evidence for the crossover is a swap in macro success rates: VP1 reaches 34.4% versus AP1's 28.2%, while AP8 reaches 39.0% versus VP8's 33.8%. The paper also claims the history benefit is genuinely temporal, not just extra conditioning capacity: replacing the ordered eight-frame history with eight copies of the current state drops action-prefix success from 39.0% to 30.8%, with a paired task-bootstrap interval excluding zero. On the 45 atomic tasks, every interface numerically beats the no-state baseline but only the discrete state prompt gives an interval-supported gain, and the best interface changes by task family.

Load-bearing premise

The load-bearing premise is that a single training seed per interface reveals the true ordering of state designs — the paper itself notes most comparisons rely on one seed — and that the 45 atomic plus 20 composite kitchen tasks stand in for robot manipulation generally.

Editorial extensions

If this is right

  • A concrete default recipe for flow-matching VLAs: put the current state frame into the vision-language side and route a short, roughly eight-frame history to the action head; anything deeper should be validated before use.
  • Long raw state histories are not a free win: past the short-window peak, added depth can degrade control, most sharply for precise knob-and-switch tasks when history is injected through the VLM prefix.
  • Interface rankings depend on task family — rearrangement favored the discrete prompt, articulation the VLM prefix, and precise actuation the state expert — so macro averages can hide where an interface helps or hurts.
  • Performance and compute separate cleanly: the discrete prompt can lead in accuracy but is by far the most expensive route, while continuous action-side interfaces add little marginal cost and scale more gently with history.
  • The evaluation protocol — fixed scaffold, independently trained systems, slot-matched controls, paired bootstrap intervals — is a reusable template for auditing how future VLAs consume robot state.

Reading between the lines

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

  • If the crossover reflects a real mechanism, interface choice may really be a temporal-horizon choice: the VLM prefix acts as a contextualizer for near-Markovian observations, while the action prefix acts as a short-term integrator; other multimodal policy architectures might show the same trade-off with observation history.
  • The appendix's cost model shows VLM-prefix attention grows quadratically with state tokens, which predicts that the VLM route becomes increasingly expensive as history lengthens — a testable extension is to hold compute fixed and see whether the crossover shifts.
  • Because the paper's state is kinematic and simulation-only, the design rule is a candidate for testing with force or tactile state and on real robots, where temporal content could matter more (contact transients) or less (noisy sensors).
  • The family-level reversals suggest future state-aware VLA reports should publish per-family success rates, since a single benchmark average can make opposite interface preferences look equivalent.
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

2 major / 4 minor

Summary. arXiv:2608.03052 studies how proprioceptive state should be fed into Vision-Language-Action policies. On a fixed flow-matching VLA (pi-0.5) with matched training data, action representation, and evaluation protocol, it compares five state interfaces—discrete state prompt, VLM prefix, action prefix, state expert, and feature modulation—against a no-state baseline on 45 atomic and 20 composite RoboCasa365 tasks, sweeping state-history depth from 1 to 96 frames. A slot-matched repeat-current control is used to separate genuine temporal content from added conditioning capacity. The paper's central claim is a design rule for its setting: with a single state frame, the state should be injected into the VLM side, while with a short multi-frame history, state should be routed to the action head.

Significance. If established, the paper would provide actionable, testable design principles for a component of VLA models that is currently wired inconsistently across systems. The strengths are real: the authors fix the backbone and training recipe, independently train each interface, reuse paired episode seeds across models, and include a slot-matched repeat-current control that directly addresses the capacity confound in history-length comparisons. The task-family partition and the analytic compute-cost model are also useful. The central limitation is that the headline routing crossover rests on single-seed point estimates, with no optimization-variance uncertainty on exactly the comparisons that drive the rule. The paper is nonetheless a careful, largely self-contained empirical study whose main claims are defensible in direction but not yet at the strength stated in the abstract.

major comments (2)
  1. The central crossover claim rests on single-seed point estimates. The paper states in Section 5.1 that paired task-bootstrap intervals 'measure evaluation and task-sampling uncertainty rather than optimization variance' and in Section 5.2 that 'most comparisons rely on a single training seed.' The decisive comparisons—VP1 34.4 vs AP1 28.2, AP8 39.0 vs VP8 33.8 on composite EEF state, and AP8 59.6 vs VP8 58.0 on atomic tasks—carry no seed-level uncertainty. Since the joint-angle panel shows AP8 and VP8 at 36.2 vs 35.8 within the paired noise band, only one panel is decisive even at the point-estimate level. A multi-seed replication of at least the four crossover cells (VP1, AP1, VP8, AP8) is needed before the abstract's testable design rule can be treated as established; alternatively, the claim should be explicitly reframed as a directional single-seed trend.
  2. The VLM-side versus action-side route comparison is confounded with interface capacity and compute path. The VLM prefix uses projector width d=2048 while the action prefix uses d=1024, and the state expert and feature modulation add 199.30M and 123.84M parameters respectively. The authors acknowledge in Section 5.2 that 'the routes differ in projection width and computation path,' but the central 'where should state enter' rule is nevertheless stated as a route preference. Because the compared systems differ in more than injection site, a matched-width sensitivity analysis or an explicit width-sweep is needed to establish that the crossover is due to the injection route itself rather than to the associated input projection and compute differences.
minor comments (4)
  1. The x-axis is labeled log scale but includes 0; consider using a linear inset or a note explaining how 0 is represented.
  2. The final row label 'AP: current-only→genuine history' is ambiguous because the columns are labeled K=1 and K=8; specify clearly that the 30.8 value is the current-only model and 39.0 is the genuine-history model.
  3. The repeat-current control is a reasonable capacity control, but it replaces ordered state with identical copies; a shuffled-order control (same frames, permuted order) would strengthen the attribution of the gain to temporal order rather than to token-repetition artifacts, and the current Appendix B evidence is descriptive rather than causal.
  4. Several words are run together in the rendered text (e.g., 'RecentVision–Language–Action(VLA)modelsalmost universally'); please ensure the camera-ready version has proper spacing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central design rule rests on controlled independent experiments and transparently disclosed empirical choices.

full rationale

Step-through of the derivation chain: (1) The interface comparisons (RQ1/RQ3) train five independently initialized systems from the same pretrained checkpoint on the same data and evaluate them closed-loop; no interface's success rate is constructed from another interface's output or from the design rule being claimed. (2) The K=8 operating point is chosen empirically from the atomic-task sweep and then applied to composite tasks; this is a recipe choice, not a fitted prediction of the composite results, and the paper explicitly calls it 'empirical rather than universal.' (3) The temporal-content claim is tested by a slot-matched repeat-current control that fixes slot count, images, language, and flow noise, comparing ordered history against repeated current state; the measured difference is exactly the temporal variation, so the inference is a controlled experiment rather than a definitional identity. (4) The compute-cost derivation in Appendix C is an analytic counting model with stated conventions (two FLOPs per multiply-add, a 4x forward factor for training), not fitted to success-rate data. (5) Self-citations such as Guo et al. 2026 appear in background motivation and related work, but no load-bearing premise, uniqueness theorem, or ansatz is imported from same-author prior work; the empirical comparisons and controls stand independently. The paper also openly discloses that most comparisons rely on a single training seed and that the rule is a directional pattern rather than an established causal mechanism; those are replicability and strength-of-evidence limitations, not circularity. No circular step can be exhibited with a quote reducing any claim to its own inputs.

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

The paper introduces no new physical entities, forces, or conservation laws. Its free parameters are design choices (history depth, quantization, widths), and its assumptions are standard domain assumptions about simulation validity and single-model representativeness. The ledger is light, which is consistent with the paper being an empirical comparison rather than a derivation.

free parameters (3)
  • history depth K=8 default = 8 frames
    Chosen empirically from the atomic-task sweep as the 'practical operating point'. The paper itself says 'This choice is empirical rather than universal', so it is a fitted operating point rather than a derived optimum.
  • state prompt quantization bins = 256 bins
    The quantization granularity for the discrete state prompt is chosen by hand and not swept; performance of the SP interface may depend on it.
  • projector widths = d=2048 for VLM prefix, d=1024 for action prefix, state expert, feature modulation
    Widths are matched to host module hidden widths, but the choice itself is a design decision; the paper notes it does not equalize parameters or compute across interfaces.
assumptions (4)
  • domain assumption RoboCasa365 simulator rollouts are a valid proxy for closed-loop robot control performance
    The entire evaluation runs in simulation; the paper's limitation section admits lack of real-robot validation.
  • domain assumption The flow-matching VLA scaffold (pi0.5) is representative of VLA models generally
    The paper studies one base policy and claims the design principles are testable guidelines for future state-aware VLAs; the generality across backbones is not tested.
  • domain assumption Task-family partition (A: rearrangement, B: articulation, C: precision) is semantically meaningful and not post hoc
    The paper states the partition is made a priori by manipulation semantics; the claim that family-level differences reflect control demands depends on this partition being stable.
  • domain assumption Paired task-bootstrap intervals measure the right uncertainty
    The paper states bootstrap intervals measure evaluation and task-sampling uncertainty, not optimization variance; most claims therefore carry no estimate of seed-to-seed training variance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Should Vision-Language-Action Models Use Proprioceptive State?." pith.science (2026). https://pith.science/paper/6IOGKYLO

@misc{pith2026260803052,
  author       = {Pith},
  title        = {Pith review of: How Should Vision-Language-Action Models Use Proprioceptive State?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IOGKYLO}},
  note         = {Machine review of arXiv:2608.03052}
}
read the original abstract

Recent Vision-Language-Action (VLA) models almost universally take robot proprioceptive state as input, yet wire it in incompatible ways -- serialized into text prompts, projected into the vision-language prefix, or fed directly to the action expert -- and almost always as a single current frame. Three questions remain open: (1) whether, and on which tasks, current state actually improves closed-loop control; (2) how much state history helps, and whether its benefit reflects genuine temporal variation rather than added conditioning capacity; and (3) where state should enter the model -- the vision-language backbone or the action-generation module. We answer these questions through controlled experiments on a flow-matching VLA, fixing the backbone, training data, action representation, and evaluation protocol throughout. We implement five representative interfaces -- discrete state prompt, VLM prefix, action prefix, state expert, and feature modulation -- under matched implementation details, and evaluate them on 45 atomic tasks spanning three task families plus 20 composite tasks; we then sweep the state-history length from 1 to 96 frames to examine how historical state information affects model performance. The experiments yield systematic answers to all three questions, distilled into testable design principles for state-aware VLAs.

Figures

Figures reproduced from arXiv: 2608.03052 by the authors.

Figure 1
Figure 1. Overview of the state design space and evaluation suite. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Estimated marginal compute of the single-frame in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Effect of state-history depth on the RoboCasa atomic subset. (a) Mean change relative to the matched single-frame [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Where and when proprioceptive state helps, and where it should enter. Macro success rate (%); dashed lines mark [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: VLM-prefix response to proprioceptive state. Layer-wise language-to-image attention redistribution and image￾token representation change are measured between matched true-state and state-off passes. VLM-side state progressively alters the shared multimodal prefix, wher…
Figure 6
Figure 6. Figure 6: State-conditioned correction over the ten-step flow trajectory. The top row reports alignment between the true￾minus-off correction and the state-off-to-expert residual. The bottom row reports correction magnitude normalized by that residual. Columns compare action-sid…
Figure 7
Figure 7. Figure 7: Representative action-conditioned spatial attention. Matched true-state and state-off passes show how a one-frame VLM prefix and one- and eight-frame action prefixes redistribute action-conditioned visual attention. The visualization is descriptive: it reveals where st…
Figure 8
Figure 8. Figure 8: Paired PrepareToast case study. (a) AP1 and AP8 are similar through placement of the two items (S1–S2), after which their reach rates separate when control must return to closing the cabinet (S3). AP8 improves S3, S4, and final success by 26 percentage points over 50 p…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 9 canonical work pages

  1. [1]

    S. Bai, M. Li, X. Lv, J. Wang, X. Wang, F. Liao, C. Hou, L.Gu,W.Zhou,K.Wu,Z.Ding,Z.Xu,L.Sun,S.Zhang, Z. Che, J. Tang, and B. Chen. HEX: humanoid-aligned experts for cross-embodiment whole-body manipulation. CoRR, abs/2604.07993,

    arXiv

  2. [4]

    Y. Dai, H. Fu, J. Lee, Y. Liu, H. Zhang, J. Yang, C. Finn, N. Fazeli, and J. Chai. Robomme: Benchmarking and un- derstanding memory for robotic generalist policies.arXiv preprint arXiv:2603.04639,

    arXiv

  3. [5]

    Davies, Y

    T. Davies, Y. Huang, Y. Liu, X. Chen, H. Liu, and L. Hu. Tenma: Robust cross-embodiment robot ma- nipulation with diffusion transformer.arXiv preprint arXiv:2509.11865,

    work page arXiv

  4. [7]

    Intelligence, K

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dha- balia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al.𝜋 0.5: a vision-language-action model with open- world generalization.arXiv preprint arXiv:2504.16054,

    arXiv

  5. [8]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakr- ishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246,

    arXiv

  6. [11]

    2602.12032

    doi: 10.48550/ARXIV. 2602.12032. URL https://doi.org/10.48550/arXiv.2602. 12032. S.Nasiriany,S.Nasiriany,A.Maddukuri,andY.Zhu. Robo- casa365: A large-scale simulation framework for train- ing and benchmarking generalist robots.arXiv preprint arXiv:2603.04356,

    work page doi:10.48550/arxiv.2602.12032

  7. [12]

    X. Qi, Y. Yang, J. Cao, L. Bai, C. Fan, C. Cao, and H. Wang. Continuous vision-language-action co-learning with semantic-physical alignment for behavioral cloning. In S. Koenig, C. Jenkins, and M. E. Taylor, editors,For- tieth AAAI Conference on Artificial Intelligence, Thirty- Eighth Conference on Innovative Applications of Arti- ficial Intelligence, Six...

    work page 2026

  8. [13]

    URL https://doi.org/10.1609/aaai.v40i29.39677

    doi: 10.1609/AAAI.V40I29.39677. URL https://doi.org/10.1609/aaai.v40i29.39677. H. Shi, B. Xie, Y. Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H.Fan,X.Zhang,andG.Huang. Memoryvla:Perceptual- cognitive memory in vision-language-action models for robotic manipulation.arXiv preprint arXiv:2508.19236,

    arXiv

Show all 24 references
  1. [14]

    F. Wang, P. Zhou, J. Qi, S. Lyu, D. Navarro-Alarcon, and G. Guo. Think proprioceptively: Embodied visual rea- soning for VLA manipulation.CoRR, abs/2602.06575,

    arXiv

  2. [15]

    URL https: //doi.org/10.48550/arXiv.2602.06575

    doi: 10.48550/ARXIV.2602.06575. URL https: //doi.org/10.48550/arXiv.2602.06575. 8 X. Wang, X. Gao, J. Fu, Z. Li, D. Fortier, G. Mullins, A. Kolobov, and B. Guo. Lola: Long horizon latent action learning for general robot manipulation.CoRR, abs/2512.20166,

    doi arXiv

  3. [16]

    doi: 10.48550/ARXIV.2512. 20166. URL https://doi.org/10.48550/arXiv.2512.20166. C.Wen,J.Lin,T.Darrell,D.Jayaraman,andY.Gao. Fight- ingcopycatagentsinbehavioralcloningfromobservation histories.AdvancesinNeuralInformationProcessingSys- tems, 33:2564–2575,

    doi

  4. [18]

    Y. Yang, M. Lin, R. Martin-Martin, M. Labrie, S. Gayaka, C.-H. Kuo, and L. Carlone. Understanding the impact of geometric foundation models on vision-language-action models.arXiv preprint arXiv:2605.24642,

    arXiv

  5. [19]

    B. Yu, S. Lian, X. Lin, Y. Wei, Z. Shen, C. Wu, Y. Miao, X. Wang, B. Wang, C. Huang, and K. Chen. Twin- brainvla: Unleashing the potential of generalist vlms for embodied tasks via asymmetric mixture-of-transformers. CoRR, abs/2601.14133,

  6. [20]

    2601.14133

    doi: 10.48550/ARXIV. 2601.14133. URL https://doi.org/10.48550/arXiv.2601. 14133. X. Zhai, B. Ou, Y. Wang, H. Y. Leong, Q. Yu, C. Hao, and Y. Liu. Cofreevla: Collision-free dual-arm manipula- tionviavision-language-actionmodelandriskestimation. arXiv preprint arXiv:2601.21712,

    doi

  7. [21]

    Zhang, J

    C. Zhang, J. Wang, Z. Gao, Y. Su, T. Dai, C. Zhou, J. Lu, and Y. Tang. CLAP: contrastive latent action pretraining for learning vision-language-action models from human videos.CoRR, abs/2601.04061,

  8. [22]

    URL https://doi.org/10

    doi: 10.48550/ARXIV.2601.04061. URL https://doi.org/10. 48550/arXiv.2601.04061. T.Z.Zhao,V.Kumar,S.Levine,andC.Finn. Learningfine- grainedbimanualmanipulationwithlow-costhardware.In K. E. Bekris, K. Hauser, S. L. Herbert, and J. Yu, editors, Robotics: Science and Systems XIX, ...

    doi

  9. [23]

    doi: 10.15607/RSS.2023. XIX.016. URL https://doi.org/10.15607/RSS.2023.XIX

    2023 doi

  10. [2019]

    W. Guo, H. Zhang, P. Li, T. Cai, Z. Chen, Y. Guo, X. He, Y.Yang,Y.Sun,andH.Xiong.Abrain-inspiredembodied intelligence for fluid and fast reflexive robotics control. arXiv preprint arXiv:2601.14628,

  11. [2020]

    Y. Wen, K. Gu, H. Liu, Y. Zhao, T. Wang, H. Fan, and X.Sun. Rosa:Harnessingrobotstatesforvision-language and action alignment.arXiv preprint arXiv:2506.13679,

    arXiv

  12. [2023]

    Whenwouldvision- proprioception policies fail in robotic manipulation? CoRR, abs/2602.12032,

    J.Lu,W.Xia,Y.Wu,Z.Lu,andD.Hu. Whenwouldvision- proprioception policies fail in robotic manipulation? CoRR, abs/2602.12032,

  13. [2024]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision- language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025a. T. Kim, J. Lee, M. Koo, D. Kim, K. Lee, C. Kim, Y. Seo, and J. Shin. Contrastive representation regularization for vision-languag...

    arXiv

  14. [2025]

    Fusai, L

    K.Black,N.Brown,D.Driess,A.Esmail,M.Equi,C.Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.𝜋0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    arXiv

  15. [2026]

    2604.07993

    doi: 10.48550/ARXIV. 2604.07993. URL https://doi.org/10.48550/arXiv.2604. 07993. J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: Anopenfoundationmodelforgeneralisthumanoidrobots. arXiv preprint arXiv:2503.14734,

    doi arXiv

  16. [4096]

    The grouped attention inner width is𝐻=8×256=2048, and policy inference uses𝑁=10 Euler denoising steps. We count one multiply-add as two floating-point opera- tions.Themarginalforwardcostofprocessingoneadditional token through all Transformer layers is approximated by 𝐹tok(𝑑,𝑚)...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.