Pith. sign in

REVIEW 2 major objections 5 minor 35 references

eDIF: A European Deep Inference Fabric for Remote Interpretability of LLM

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A European NDIF-compatible interpretability cluster proved viable in a 16-researcher pilot.

desk verdict An honest feasibility report on deploying NDIF-style interpretability infrastructure in Europe; the technical claims hold up, but the user-satisfaction numbers should not be quoted without response counts. read the letter →

arxiv 2508.10553 v1 pith:GCVWZ3GV submitted 2025-08-14 cs.CL

Irma Heithoff. Marc Guggenberger , Sandra Kalogiannis , Susanne Mayer , Fabian Maag , Sigurd Schacht , Carsten Lanquillon This is my paper
classification cs.CL
keywords mechanisticinterpretabilitylargelanguagemodelsNDIFNNsightremoteinferenceinfrastructurefeasibilitystudyactivationpatchingGPUcluster
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 feasibility study claims that a European, NDIF-compatible interpretability cluster (eDIF) can be built on ordinary institutional GPU hardware and that remote mechanistic-interpretability experiments are practical and well received. The authors deployed an eight-GPU cluster, exposed it through the NNsight deferred-execution API, and ran a six-week pilot with 16 academic researchers across seven countries. Participants performed activation patching, causal tracing, and representation analysis on models up to DeepSeek-R1-70B; usage climbed to over 700 requests in a single peak day while the system stayed stable. The paper concludes that shared remote interpretability infrastructure is viable in Europe, that users strongly support sustained access, and that the main barriers are bandwidth, model-switching, and error handling rather than fundamental design flaws.

What carries the argument

The load-bearing mechanism is NNsight's deferred execution model: researchers write standard PyTorch-style trace code locally that specifies interventions (e.g., patching a Transformer activation), and a FastAPI/Ray backend executes those traces remotely on shared, pre-loaded model instances running in Docker containers on GPU servers. This decouples experiment design from model execution, enabling co-tenancy—multiple users sharing pooled GPUs—and remote access to large models without local deployment. The paper's portability result hinges on PyTorch's runtime abstraction: moving from NVIDIA/CUDA to AMD/ROCm required only replacing NCCL with RCCL, and moving from Docker to Apptainer for HPC

What would settle it

A replication study that requires all 16 participants to complete a post-task questionnaire—with response counts, per-item variance, and independent error logs—would settle whether the 4.2/5 onboarding and 4.0/5 intuitiveness ratings, and the unanimity on sustained access, are robust; likewise, a public trace of GPU utilization during the peak 700-request day would confirm the claimed stability.

Watch

Extended reading notes

Core claim

The central discovery is that a NDIF-style remote interpretability deployment is feasible outside its original U.S. context and can support real research workflows. The eDIF cluster, hosted at a German university of applied sciences with eight RTX A6000 GPUs, ran GPT-2, DeepSeek-R1-Distill-Llama-8B, and DeepSeek-R1-Distill-Llama-70B behind the NNsight API. In the pilot, researchers executed deferred interventions remotely, evaluated the API as superior to local toolkits for this purpose, and reported mean onboarding/ease-of-use ratings of 4.2/5 and intuitiveness 4.0/5, with 83.3% partially integrating the platform into their research and 100% requesting continued access. The paper also shows

Load-bearing premise

The user-satisfaction conclusions rest on feedback from a self-selected subset of the 16 pilot participants, and the paper gives no response counts or variance; if those respondents are unrepresentative of the broader research community, the positive reception and 100% sustained-access demand may not hold.

Editorial extensions

If this is right

  • European academic groups without local HPC can run activation-level interventions on 70B-parameter models, lowering the entry barrier to mechanistic interpretability.
  • The reported 100% demand for sustained access and 83.3% partial workflow integration suggest a durable user community is forming around shared interpretability infrastructure.
  • Feasible AMD/ROCm porting means future scaling can use cheaper GPU options without being locked to a single vendor.
  • The identified pain points—30-minute activation downloads, intermittent freezes, lack of model hot-swapping—define a concrete engineering roadmap for production readiness.
  • If no-code/low-code tooling matches the reported demand, interpretability methods could extend beyond specialist programmers to students and interdisciplinary collaborators.

Reading between the lines

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

  • An untested implication: the 'time savings' estimates of 10–50% rest on self-report; a controlled comparison against local execution with equivalent hardware could quantify real productivity gains.
  • If GPU-sharing inefficiency (unused memory on per-model allocation) is addressed with automatic profiling, the same eight-GPU cluster could plausibly serve a larger user base at marginal cost.
  • The pilot's recruitment via Discord and e-mail likely over-selects technically sophisticated users, so the reported ease-of-use scores may not generalize to the broader ML community; a less self-selected cohort would be a sterner test.
  • A natural next experiment is a multi-national deployment bridging two or more eDIF nodes with different GPUs, which would test the governance and cross-institutional scheduling questions the current study leaves open.
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 / 5 minor

Summary. The paper reports a feasibility study of eDIF, a European NDIF-compatible remote interpretability cluster. It describes the system architecture (eight NVIDIA RTX A6000 GPUs, Docker/Ray/FastAPI, monitoring stack), model provisioning (GPT-2, DeepSeek-R1-Distill-Llama-8B/70B), a six-week pilot with 16 selected European researchers, usage and error data, NNsight-based remote interpretability experiments, early experiences porting the stack to AMD ROCm on an HPC cluster, user feedback, and a roadmap. The central claim is that eDIF demonstrates the viability and research value of such infrastructure in Europe, with the abstract and conclusion asserting positive reception and confirmatory user evaluations.

Significance. The paper makes a useful empirical contribution to the infrastructure literature on shared LLM interpretability. Its strengths are the concrete operational evidence: a real six-week deployment, peak usage above 700 daily requests, deployment at the scale of DeepSeek-R1-70B, and a documented CUDA-to-ROCm migration on an HPC cluster. The authors also report operational problems candidly (30-minute activation downloads, intermittent freezes, multi-GPU shared-parameter crashes). The main weakness is that the survey-based evidence for the 'positive reception' and 'research value' claims is underreported: response counts, variance, and exact denominators are missing, and Section VII.B itself notes that feedback participation was low. The technical feasibility conclusion is supported by monitoring data, but the user-evaluation half of the headline claim needs either additional reporting or tempering before the paper can stand as written.

major comments (2)
  1. [§VI.B, §VII.B, Abstract, §IX] The user-feedback section reports means of 4.2/5 and 4.0/5, '83.3%' integration, '83.4%' support for community features, and '100%' support for sustained access without any response counts, per-item distributions, or variance. Section VII.B explicitly admits that 'feedback participation was relatively low.' Section IX nevertheless concludes that 'positive evaluations ... confirm the platform's relevance,' and the abstract claims 'positive reception.' These conclusions exceed the reported evidence. Please report N for every percentage and mean, the exact response rate, the distribution of responses, and a statement of how the self-selected subset limits generalizability; alternatively, rephrase the abstract and conclusion to say 'preliminary positive feedback from a subset of pilot participants.' This is a reporting fix and does not undermine the technical feasibility evidence.
  2. [§V, §VI.B, §IX] The pilot population is a self-selected applicant pool (29 applicants) from which 16 researchers were chosen on the basis of expertise and planned experiments; the feedback respondents are a further unreported subset. The conclusion's phrase 'reproducible experimentation for a diverse group of researchers' is not the issue, but broader statements about the general research community are. Please define the denominator for each feedback item and note explicitly that the reported satisfaction scores reflect only those who chose to respond. If the full questionnaire was administered to all 16, state the response count; if it was not, say so. This is necessary to make the user-experience claims falsifiable.
minor comments (5)
  1. [Abstract, §VI.A] 'Stable platform performance throughout' is stronger than the reported intermittent freezes and error states in Figure 4. Please quantify the error rate (e.g., percentage of requests returning errors) or soften the wording to 'stable enough for pilot experimentation.'
  2. [§VIII.A] The sentence 'The current server servers with RTX 6000 and H100 at Ansbach University of Applied Sciences, equipped with NVIDIA RTX A6000 GPUs...' is garbled. Clarify which hardware remains in operation.
  3. [§VIII.B] The paragraph beginning 'In parallel, online community spaces (e.g., Discord, GitHub)...' is duplicated verbatim after the first paragraph of Section VIII.B.
  4. [§V, §I] Typo: 'infrastructur' should be 'infrastructure.' Also, the sentence 'A scalable, open infrastructure for mechanistic interpretability research based in Europe. In cooperation...' is a fragment; consider joining it to the following sentence.
  5. [Figures 3 and 4] Both figures need axis labels, units, and captions that define what counts as a 'request' and what counts as an 'error.' Without this, the reader cannot interpret the 700-request peak or the error comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical infrastructure feasibility report whose central claims rest on deployment, monitoring, and usage data, not on fitted parameters or self-referential derivations.

full rationale

This paper is a feasibility study of a computing infrastructure deployment. It does not claim to derive a prediction from first principles, nor does it fit parameters and then rename those fits as predictions. The central claims—that NDIF-compatible infrastructure can be deployed in Europe, that models such as GPT-2 and DeepSeek-R1-70B can be served remotely, and that users can perform interventions via NNsight—are supported by operational evidence: actual GPU provisioning, request-volume monitoring (Figure 3, over 700 requests/day), error logs (Figure 4), and model deployment tables. The conclusion's 'positive evaluations' component rests on user feedback, and the paper itself flags in Section VII.B that 'feedback participation was relatively low, potentially limiting the comprehensiveness of insights and generalizability of the findings.' That is a stated evidentiary limitation, not circularity: the feedback is not an input that is then repackaged as the derivation's output; it is simply a measurement with known sampling weaknesses. The collaboration with and citation of the NDIF team is external and not a self-citation chain that forces the paper's conclusions. No equation, definition, or fitted quantity is shown to be equivalent by construction to a claimed result. Therefore the circularity score is 0.

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

The paper rests on the correctness of the NDIF/NNsight stack and on the representativeness of self-reported user feedback. No free parameters or invented entities appear: this is an infrastructure and user-study report, not a fitting or derivation exercise.

assumptions (2)
  • domain assumption The NDIF stack and NNsight API behave as documented in the cited guide and papers.
    The entire deployment follows the NDIF development guide (Section IV.A) and assumes its correctness; any flaw in NDIF would affect eDIF results.
  • domain assumption Self-reported user feedback is a valid proxy for scientific utility.
    Sections VI.B and VII.B use user ratings and time-saving estimates as evidence of the platform's value, without independent verification or response-rate reporting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of eDIF: A European Deep Inference Fabric for Remote Interpretability of LLM." pith.science (2026). https://pith.science/paper/GCVWZ3GV

@misc{pith2026250810553,
  author       = {Pith},
  title        = {Pith review of: eDIF: A European Deep Inference Fabric for Remote Interpretability of LLM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GCVWZ3GV}},
  note         = {Machine review of arXiv:2508.10553}
}
read the original abstract

This paper presents a feasibility study on the deployment of a European Deep Inference Fabric (eDIF), an NDIF-compatible infrastructure designed to support mechanistic interpretability research on large language models. The need for widespread accessibility of LLM interpretability infrastructure in Europe drives this initiative to democratize advanced model analysis capabilities for the research community. The project introduces a GPU-based cluster hosted at Ansbach University of Applied Sciences and interconnected with partner institutions, enabling remote model inspection via the NNsight API. A structured pilot study involving 16 researchers from across Europe evaluated the platform's technical performance, usability, and scientific utility. Users conducted interventions such as activation patching, causal tracing, and representation analysis on models including GPT-2 and DeepSeek-R1-70B. The study revealed a gradual increase in user engagement, stable platform performance throughout, and a positive reception of the remote experimentation capabilities. It also marked the starting point for building a user community around the platform. Identified limitations such as prolonged download durations for activation data as well as intermittent execution interruptions are addressed in the roadmap for future development. This initiative marks a significant step towards widespread accessibility of LLM interpretability infrastructure in Europe and lays the groundwork for broader deployment, expanded tooling, and sustained community collaboration in mechanistic interpretability research.

Figures

Figures reproduced from arXiv: 2508.10553 by the authors.

Figure 1
Figure 1. Layered architecture of the eDIF system with user interaction, orchestration, and GPU execution FAU HPC environment required porting these core services to Apptainer (formerly Singularity), a container runtime compat￾ible with non-root, multi-user HPC systems. Docker-specific features such as image builds, service orchestration, and net￾working had to be either manually reconfigured or replaced with lightweight shel… view at source ↗
Figure 2
Figure 2. NNsight distributed execution architecture for remote neural network inspection GPU allocation for each LLM was based on expected mem￾ory requirements. This approach provided reasonable resource planning but created inefficiencies since GPUs could not be shared between models. Unused memory on individual GPUs remained inaccessible to other models, creating a gap between theoretical GPU capacity and actual utilizatio… view at source ↗
Figure 3
Figure 3. eDIF server activities during feasibility study [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: eDIF requests – Error comparison during feasibility study User experience evaluations reflected high satisfaction, with mean ratings of 4.2 out of 5 for onboarding and ease of use, and 4.0 for the intuitiveness of the overall framework. Participants particularly emphas…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages

  1. [1]

    TransformerLens,

    N. Nanda and J. Bloom, “TransformerLens,” GitHub reposi- tory, 2022. [Online]. Available: https://github.com/TransformerLensOrg/ TransformerLens

    work page 2022

  2. [2]

    An explainable transformer circuit for compositional generalization,

    C. Tang, B. Lake, and M. Jazayeri, “An explainable transformer circuit for compositional generalization,” arXiv preprint arXiv:2502.15801 , Feb. 2025. [Online]. Available: http://arxiv.org/abs/2502.15801

    arXiv 2025

  3. [3]

    Mechanistic interpretability for AI safety – A review,

    L. Bereska and E. Gavves, “Mechanistic interpretability for AI safety – A review,” arXiv preprint arXiv:2404.14082 , Apr. 2024. [Online]. Available: http://arxiv.org/abs/2404.14082

    arXiv 2024

  4. [4]

    The 2024 foundation model transparency index,

    R. Bommasani, K. Klyman, S. Kapoor, S. Longpre, B. Xiong, N. Maslej, and P. Liang, “The 2024 foundation model transparency index,” arXiv preprint arXiv:2407.12929, Mar. 2025. [Online]. Available: http://arxiv. org/abs/2407.12929

    arXiv 2024

  5. [5]

    Comprehensive analysis of transparency and accessibility of ChatGPT, DeepSeek, and other SoTA large language models,

    R. Sapkota, S. Raza, and M. Karkee, “Comprehensive analysis of transparency and accessibility of ChatGPT, DeepSeek, and other SoTA large language models,” arXiv preprint arXiv:2502.18505 , Feb. 2025. [Online]. Available: http://arxiv.org/abs/2502.18505

    arXiv 2025

  6. [6]

    NNsight and NDIF: Democratizing ac- cess to open-weight foundation model internals,

    J. Fiotto-Kaufman et al. , “NNsight and NDIF: Democratizing ac- cess to open-weight foundation model internals,” arXiv preprint arXiv:2407.14561, Apr. 2025. [Online]. Available: http://arxiv.org/abs/ 2407.14561

    arXiv 2025

  7. [7]

    pyvene: A library for understand- ing and improving PyTorch models via interventions,

    Z. Wu, A. Geiger, A. Arora, J. Huang, Z. Wang, N. D. Goodman, C. D. Manning, and C. Potts, “pyvene: A library for understand- ing and improving PyTorch models via interventions,” arXiv preprint arXiv:2403.07809, Mar. 2024. [Online]. Available: http://arxiv.org/abs/ 2403.07809

    arXiv 2024

  8. [8]

    baukit: Tools for inspecting and intervening in PyTorch models,

    D. Bau, “baukit: Tools for inspecting and intervening in PyTorch models,” GitHub repository, 2022. [Online]. Available: https://github. com/davidbau/baukit

    work page 2022

Show all 35 references
  1. [9]

    LUMI: One of the EuroHPC world-class su- percomputers,

    LUMI Supercomputer, “LUMI: One of the EuroHPC world-class su- percomputers,” [Online]. Available: https://lumi-supercomputer.eu [Ac- cessed: Jul. 18, 2025]

    2025

  2. [10]

    MareNostrum 5,

    Barcelona Supercomputing Center, “MareNostrum 5,” [Online]. Avail- able: https://www.bsc.es/marenostrum/marenostrum-5 [Accessed: Jul. 18, 2025]

    2025

  3. [11]

    NVIDIA RTX A6000 Datasheet,

    NVIDIA, “NVIDIA RTX A6000 Datasheet,” 2023. [Online]. Avail- able: https://www.nvidia.com/en-us/design-visualization/rtx-a6000/ [Ac- cessed: Jul. 18, 2025]

    2023

  4. [12]

    Ubuntu 22.04 LTS released,

    Canonical, “Ubuntu 22.04 LTS released,” Apr. 21, 2022. [Online]. Avail- able: https://canonical.com/blog/ubuntu-22-04-lts-released [Accessed: Jul. 18, 2025]

    2022

  5. [13]

    Scalable parallel programming with CUDA,

    J. Nickolls, I. Buck, M. Garland, and K. Skadron, “Scalable parallel programming with CUDA,” ACM Queue, vol. 6, no. 2, pp. 40–53, 2008. DOI: 10.1145/1365490.1365500

    2008

  6. [14]

    Overview — Ray,

    Anyscale Inc., “Overview — Ray,” [Online]. Available: https://docs.ray. io/en/latest/ray-overview/index.html [Accessed: Jul. 12, 2025]

    2025

  7. [15]

    MinIO: High-performance, S3 compatible object storage,

    MinIO Inc., “MinIO: High-performance, S3 compatible object storage,” [Online]. Available: https://min.io/ [Accessed: Jul. 12, 2025]

    2025

  8. [16]

    Grafana Loki

    Grafana, “Grafana Loki” [Online]. Available: https://grafana.com/oss/ loki/ [Accessed: Jul. 22, 2025]

    2025

  9. [17]

    Prometheus: Overview,

    Prometheus, “Prometheus: Overview,” [Online]. Available: https:// prometheus.io/docs/introduction/overview/ [Accessed: Jul. 22, 2025]

    2025

  10. [18]

    InfluxDB Overview

    influxdata, “InfluxDB Overview” [Online]. Available: https://www. influxdata.com/products/influxdb-overview/ [Accessed: Jul. 22, 2025]

    2025

  11. [19]

    Available: https://grafana.com/grafana/ [Accessed: Jul

    Grafana, “Grafana” [Online]. Available: https://grafana.com/grafana/ [Accessed: Jul. 22, 2025]

    2025

  12. [20]

    NCCL – NVIDIA Collective Communications Library,

    NVIDIA Corporation, “NCCL – NVIDIA Collective Communications Library,” [Online]. Available: https://developer.nvidia.com/nccl [Ac- cessed: Aug. 3, 2025]

    2025

  13. [21]

    RCCL – ROCm Collective Communica- tion Library,

    Advanced Micro Devices, Inc., “RCCL – ROCm Collective Communica- tion Library,” [Online]. Available: https://rocm.docs.amd.com/projects/ rccl [Accessed: Aug. 3, 2025]

    2025

  14. [22]

    NDIF: Development Guide,

    J. Fiotto-Kaufman et al., NDIF, “NDIF: Development Guide,” [Online]. Available: https://github.com/ndif-team/ndif [Accessed: Jul. 12, 2025]

    2025

  15. [23]

    Available: https://huggingface.co/ openai-community/gpt2 [Accessed: Jul

    Hugging Face, “GPT-2,” [Online]. Available: https://huggingface.co/ openai-community/gpt2 [Accessed: Jul. 22, 2025]

    2025

  16. [24]

    DeepSeek-R1-Distill-Llama-8B,

    Hugging Face, “DeepSeek-R1-Distill-Llama-8B,” [Online]. Available: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B [Accessed: Jul. 22, 2025]

    2025

  17. [25]

    DeepSeek-R1-Distill-Llama-70B,

    Hugging Face, “DeepSeek-R1-Distill-Llama-70B,” [Online]. Available: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B [Accessed: Jul. 22, 2025]

    2025

  18. [26]

    Singularity: Scientific containers for mobility of compute,

    Kurtzer, G.M., Sochat, V . and Bauer, M.W., “Singularity: Scientific containers for mobility of compute,” *PLOS ONE*, 2017. [Online]. Available: https://doi.org/10.1371/journal.pone.0177459 [Accessed: Jul. 22, 2025]

    2017 doi

  19. [27]

    How to use and inter- pret activation patching,

    Heimersheim, Stefan and Nanda, Neel, “How to use and inter- pret activation patching,” [Online]. Available: https://arxiv.org/abs/2404. 15255, doi: https://doi.org/10.48550/arXiv.2404.15255 [Accessed: Jul. 20, 2025]

    2025 doi arXiv

  20. [28]

    interpreting GPT: the logit lens,

    nostalgebraist, “interpreting GPT: the logit lens,” [Online]. Available: https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/ interpreting-gpt-the-logit-lens [Accessed: Jul. 20, 2025]

    2025

  21. [29]

    Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models,

    Marks, Samuel; Rager, Can; Michaud, Eric J.; Belinkov, Yonatan; Bau, David; Mueller, Aaron, “Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models,” CoRR, vol.abs/2403.19647, 2024, [Online]. Available: https://arxiv.org/abs/ 2403.19647...

    2024 doi arXiv

  22. [30]

    LLM Circuit Analyses Are Consistent Across Training and Scale,

    Tigges, Curt; Hanna, Michael; Yu, Qinan; Biderman, Stella, “LLM Circuit Analyses Are Consistent Across Training and Scale,” CoRR, vol. abs/2407.10827, 2024, [Online]. Available: https://arxiv.org/abs/2407. 10827, doi: https://doi.org/10.48550/arXiv.2407.10827 [Accessed: Jul. 20, 2025]

    2024 doi arXiv

  23. [31]

    Available: https://github.com/ QwenLM/Qwen [Accessed: Jul

    Alibaba DAMO, “QwenLM,” [Online]. Available: https://github.com/ QwenLM/Qwen [Accessed: Jul. 20, 2025]

    2025

  24. [32]

    Available: https://deepmind

    Google DeepMind, “Gemma,” [Online]. Available: https://deepmind. google/models/gemma/ [Accessed: Jul. 20, 2025]

    2025

  25. [33]

    A Mathematical Framework for Transformer Circuits,

    Nelson Elhage, Neel Nanda, Tom Henighan, Chris Olah, Others, “A Mathematical Framework for Transformer Circuits,” [Online]. Avail- able: https://transformer-circuits.pub/2021/framework/index.html [Ac- cessed: Jul. 20, 2025]

    2021

  26. [34]

    Neuronpedia,

    www.neuronpedia.org, “Neuronpedia,” [Online]. Available: https://www. neuronpedia.org [Accessed: Jul. 20, 2025]

    2025

  27. [35]

    AMD vs NVIDIA Inference Benchmark: Who Wins? – Performance & Cost Per Million Tokens,

    www.semianalysis.com, “AMD vs NVIDIA Inference Benchmark: Who Wins? – Performance & Cost Per Million Tokens,” [Online]. Available: https://semianalysis.com/2025/05/23/ amd-vs-nvidia-inference-benchmark-who-wins-performance-cost-per-million-tokens/ [Accessed: Jul. 27, 2025]. 9

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.