Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Analyzing Character Representation in Media Content using Multimodal Foundation Model: Effectiveness and Trust

T0 review · 3 major / 6 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 The paper's user study finds that lay audiences can read AI-generated gender and age representation charts and rate the tool useful, but trust in the age model is only slight.

desk verdict Real first step on audience trust in AI representation analytics, but the abstract overstates the understanding result. read the letter →

arxiv 2506.14799 v2 pith:Z732FFHT submitted 2025-06-02 cs.HC cs.AIcs.LG

Evdoxia Taka , Debadyuti Bhattacharya , Joanne Garde-Hansen , Sanjay Sharma , Tanaya Guha This is my paper
classification cs.HCcs.AIcs.LG
keywords multimodalfoundationmodelCLIPcharacterrepresentationgenderandageanalysisvisualizationAItrustuserstudymediacontent
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 builds a tool that automatically measures how often male/female and older/younger faces appear on screen in full-length films, using the CLIP vision-language model, and then asks ordinary viewers to interpret the results. Its central empirical claim is that, with a carefully designed visualization, lay audiences can correctly read gender and age representation statistics, find the tool moderately useful, and value seeing the AI's confidence and bias estimates. The same study finds that trust is only moderate for gender predictions and lower for age predictions. If these findings hold, automated representation analytics can be put in front of audiences rather than reserved for researchers, with the caveat that trust and comprehension are not automatic.

What carries the argument

The pipeline begins with a Single Shot Scale-Invariant Face Detector on video frames, extracts CLIP image embeddings for each detected face, and feeds those embeddings into logistic regression classifiers trained and benchmarked on the FairFace dataset to predict perceived gender (female/male) and age group (nine bands from 0–2 to 70+). The visualization maps age to 'up to 50' vs 'over 50' and gender at an intersection using a doughnut chart with a nested ring and hover popups, and adds bar graphs showing model confidence and bias. The user-study analysis uses Bayesian intervals to estimate the probability of correct answers and mean Likert ratings, so conclusions are stated as posterior means and 94% credible intervals.

What would settle it

A replication with a demographically representative sample (for instance, several hundred participants matched to national age, education, and media-use distributions) that found no better-than-chance performance on Q2.1 or substantially lower usefulness ratings would undercut the paper's central claim. A simpler direct check would be to re-analyze the released study data to test whether comprehension and trust correlate with participant age or AI familiarity.

Watch

Extended reading notes

Core claim

The paper's central claim is that a CLIP-based analysis pipeline plus a nested doughnut-chart visualization can communicate AI-generated demographic representation statistics to a general audience well enough to be useful, while making visible the limits of the models. In a 30-participant user study, most participants correctly answered questions comparing gender and age prevalence across films (posterior mean probability around 0.86–0.87) and correctly identified which film favored women over 50 and where the AI would under-detect. Self-reported trust averaged 3.29 on a 5-point scale for gender predictions and 2.58 for age predictions, i.e. moderate and slightly-below-moderate respectively. Participants rated the tool's overall usefulness moderately, while rating the AI confidence and bias information as useful, and several asked for more demographic categories and contextual information about characters.

Load-bearing premise

The study's conclusions about how the general public understands and trusts the tool rest on 30 participants recruited through university mailing lists and social media, 86.67% of whom were 40 or younger; if that group is more educated and more AI-familiar than the broader public, the observed comprehension and usefulness ratings may not generalize.

Editorial extensions

If this is right

  • Lay audiences can use well-designed charts to compare which gender or age group dominates screen time in a film.
  • Presenting model confidence and bias alongside results does not reduce self-reported usefulness; participants rated this information as useful.
  • Designers of such tools should expect age predictions to be trusted less than gender predictions and should therefore provide explanations or uncertainty cues for age.
  • Audiences want more than numbers: additional demographics (e.g., ethnicity) and contextual information about how characters are portrayed.
  • The tool can support media-literacy applications, such as analyzing temporal trends or comparing films across genres, without requiring expertise in computational media analysis.

Reading between the lines

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

  • Extending the unit of analysis from on-screen face appearances to character-level counts would directly address the Q2.1 confusion about what the percentages count; the current data cannot tell whether the ambiguity is a wording problem or a conceptual one.
  • The lower trust in age predictions suggests that showing confidence scores alone may not be enough; a tool that displays concrete failure cases (e.g., misclassified age groups) would test whether transparency or model accuracy drives trust.
  • Applied to streaming or social video, the same pipeline could generate per-scene or per-decade representation timelines, letting users connect numbers to narrative context instead of comparing whole films only.
  • Given the sample skew toward younger participants, re-weighting the released user-study data by population age distributions would give a rough first estimate of how the headline trust and usefulness numbers might shift for the general public.
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

3 major / 6 minor

Summary. The paper presents a CLIP-based pipeline that detects faces in films, predicts perceived gender and age from face CLIP embeddings, and visualizes on-screen appearance distributions in doughnut charts aimed at lay audiences. The authors benchmark the gender/age classifiers on FairFace, extract analytics from three films, and run a 30-participant user study to assess understanding (RQ1), trust in the AI models (RQ2), usefulness (RQ3), and user experience (RQ4). The results show high accuracy on simple perceptual read-offs (Q2.2–Q2.3), moderate understanding of the underlying semantics of the displayed percentages (Q2.1, with a 94% CI of 0.47–0.78), moderate trust in gender predictions and lower trust in age predictions, and moderate-to-high perceived usefulness of the tool. The paper concludes that participants understood the analytics and found the tool overall useful, while expressing a need for more demographic categories and contextual character information.

Significance. If its claims hold, this work fills a real gap: earlier AI-based media representation analyses stopped at extracting statistics, without asking whether and how audiences understand and trust them. The paper's strengths include its open-source release of code, benchmarks, and user-study data; the use of Bayesian analysis appropriate for a small sample; and an unusually candid discussion that acknowledges the ambiguity of several key results (e.g., Q2.1 and Q2.4). The finding that trust in the age model is lower than in the gender model, and that users found AI confidence and bias information useful, is a useful empirical contribution for designers of media-analytics interfaces. However, the headline claim about user understanding is not fully supported by the paper's own primary comprehension item, which limits the strength of the central 'effectiveness' contribution.

major comments (3)
  1. [Abstract and Section 4.3] The claim in the abstract and Section 5 that 'participants were able to understand the analytics in our visualizations' is not supported by the paper's main comprehension item, Q2.1, which asks what the displayed percentages represent. The reported 94% Bayesian CI for correct response is 0.47–0.78 (mean 0.62), straddling the 0.5 chance line, and only 19 of 30 participants answered correctly. The paper itself concedes 'the evidence on the ability of the participants to understand the character distribution ... is not strong (Q2.1)'. Because RQ1 is central to the stated contribution, the conclusions must be tempered, or the authors should provide a composite analysis showing that understanding is established across a pre-registered set of items despite Q2.1.
  2. [Section 3.2] The binarization of age predictions into 'Up to 50' and 'Over 50' is never validated. The rule assigns a face to 'Over 50' if the summed confidences of the >50 age classes exceed the summed confidences of the <50 classes, but the benchmarked classifier uses the FairFace nine-class scheme, where age 50 falls inside the '50-59' class; the handling of age 50 is unspecified. Moreover, the accuracy of the binarized classifier could differ substantially from the 9-class accuracy in Table 1, and all age-related conclusions (Q2.3, Q2.5, Q2.6) depend on this binarization. Please report the binarized model's accuracy/F1 and state explicitly how the '50-59' class is assigned.
  3. [Section 4.1 and Figure 3] The study sample is not representative of the 'general public' or 'lay audience' that the paper targets. All 30 participants were recruited through University of Glasgow mailing lists and social media channels, and 86.67% of them were aged 40 or under (Figure 3). Because trust in AI and interpretation of visualizations are likely to vary with age and technical background, the current sample cannot support generalizing conclusions about what 'the audience' understands or trusts. The discussion should explicitly address this limitation and soften any claims about broader population conclusions.
minor comments (6)
  1. [Abstract] Minor grammar issue: 'Recent advances in AI has made' should be 'Recent advances in AI have made'.
  2. [Section 3.2] The description of the age-group assignment would benefit from a precise definition of how the boundary at age 50 is handled, since the underlying classes are 40-49, 50-59, 60-69, and 70+.
  3. [Section 4.3] The sentence 'CIs that are on the right of the reference line of 0.5 and do not intersect with it indicate the probability of providing accurate responses' is unclear; suggest rephrasing to 'indicate a probability of accurate responses significantly above chance'.
  4. [Table 3] The response options for Q2.1 are not clearly separated typographically; separating them with semicolons or line breaks would improve readability.
  5. [Section 4.3 (Q2.4)] The scoring of Q2.4 is ambiguous because the authors themselves identify two defensible correct answers ('Female Over 50' and 'Male Over 50'). Reporting a CI for this item without either excluding it or re-scoring it risks misleading readers; a sensitivity analysis would be helpful.
  6. [Section 3.2] The label 'bias' for the bar graphs that show actual versus predicted percentages in the validation set is potentially misleading; this is better described as a calibration display, and the term 'bias' should be defined precisely in the text to avoid confounding the trust measurements.

Circularity Check

0 steps flagged · score 0.0 of 10

No meaningful circularity; the paper's central claims rest on direct user-study measurements rather than on any model-derived prediction.

full rationale

The paper's main claims—that participants understood the analytics, found the tool moderately useful, and trusted gender predictions more than age predictions—are established by direct questionnaire responses analyzed with standard Bayesian interval estimation. There is no fitted parameter that is later renamed a prediction, and no quantity used as evidence is defined in terms of the conclusion it supports. The analytics extraction pipeline is benchmarked against the external FairFace dataset, and the user study is independent of the model outputs in the sense that it measures self-reported comprehension and trust rather than deriving them from the CLIP classifiers. The paper self-cites prior work (e.g., [28] for the Bayesian analysis method and [11, 15, 26] for related media-analysis frameworks), but these citations are methodological or contextual and do not carry the load of the empirical conclusions; the Bayesian method is standard and externally established. The acknowledged weakness of Q2.1, whose 94% CI crosses 0.5, is a limitation on the strength of the evidence and a possible overstatement in the abstract, not a circularity. Likewise, showing model bias computed on the FairFace validation set is a design choice that may affect trust judgments but does not make the user-study outcome equivalent to the pipeline's inputs. Overall, the derivation chain is self-contained: data are collected from participants, analyzed statistically, and reported as empirical findings.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper's central claims rest on several assumptions: FairFace labels are a valid ground truth for perceived demographics in films, CLIP embeddings encode enough demographic signal, the bias statistics on FairFace transfer to the films, and the user-study sample represents the general audience. The only fitted decision rule is the age binarization threshold, which is a modeling choice rather than a tuned parameter. No new entities such as particles, mediators, or forces are introduced.

free parameters (1)
  • Age binarization rule = 50/50 soft confidence split
    The 9 FairFace age classes are mapped to 'Up to 50' vs 'Over 50' by whichever sum of softmax confidences is larger; this is a modeling choice not separately validated on the binarized task and it directly determines all reported film age statistics (Section 3.2).
assumptions (4)
  • domain assumption FairFace perceived age and gender labels are an appropriate ground truth for character representation in films.
    The classifiers are trained and benchmarked on FairFace, and the visualizations use these labels as the categories of interest (Section 3.1).
  • domain assumption CLIP face embeddings encode enough demographic information for logistic regression to achieve near-state-of-the-art gender and age classification.
    The pipeline depends on CLIP embeddings as the sole face representation for the classifiers (Section 3.1).
  • domain assumption Model bias statistics computed on the FairFace validation set are informative about model behavior on the analyzed films.
    The visualization presents actual vs predicted FairFace percentages as 'bias' information to users, implicitly assuming transferability to the films, for which no ground-truth labels exist (Section 3.2).
  • domain assumption The 30 university-recruited participants represent the general audience.
    The motivating question asks about usefulness and trust for the general public, but recruitment was through University of Glasgow mailing lists and social media channels (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analyzing Character Representation in Media Content using Multimodal Foundation Model: Effectiveness and Trust." pith.science (2026). https://pith.science/paper/Z732FFHT

@misc{pith2026250614799,
  author       = {Pith},
  title        = {Pith review of: Analyzing Character Representation in Media Content using Multimodal Foundation Model: Effectiveness and Trust},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z732FFHT}},
  note         = {Machine review of arXiv:2506.14799}
}
read the original abstract

Recent advances in AI has made automated analysis of complex media content at scale possible while generating actionable insights regarding character representation along such dimensions as gender and age. Past works focused on quantifying representation from audio/video/text using AI models, but without having the audience in the loop. We ask, even if character distribution along demographic dimensions are available, how useful are those to the general public? Do they actually trust the numbers generated by AI models? Our work addresses these open questions by proposing a new AI-based character representation tool and performing a thorough user study. Our tool has two components: (i) An analytics extraction model based on the Contrastive Language Image Pretraining (CLIP) foundation model that analyzes visual screen data to quantify character representation across age and gender; (ii) A visualization component effectively designed for presenting the analytics to lay audience. The user study seeks empirical evidence on the usefulness and trustworthiness of the AI-generated results for carefully chosen movies presented in the form of our visualizations. We found that participants were able to understand the analytics in our visualizations, and deemed the tool `overall useful'. Participants also indicated a need for more detailed visualizations to include more demographic categories and contextual information of the characters. Participants' trust in AI-based gender and age models is seen to be moderate to low, although they were not against the use of AI in this context. Our tool including code, benchmarking, and the user study data can be found at https://github.com/debadyuti0510/Character-Representation-Media.

Figures

Figures reproduced from arXiv: 2506.14799 by the authors.

Figure 1
Figure 1. Our tool to analyze the demographic character representation in videos. It detects the faces in the video, generates the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Charts shown to participants with character repre [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Part 1: pre-questionnaire & distribution of responses. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Part 2-3: 94% Bayesian CIs for (a) the probability of correct response in the information [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 22 canonical work pages

  1. [1]

    David Bamman, Rachael Samberg, Richard Jean So, and Naitian Zhou. 2024. Mea- suring diversity in Hollywood through the large-scale computational analysis of film. Proceedings of the National Academy of Sciences 121, 46 (2024), e2409770121

    work page 2024

  2. [2]

    Centre for Ageing Better. 2023. Cast aside: Exploring the presence of older characters in British films. https://ageing-better.org.uk/sites/default/files/2023- 02/cast-aside-insight-report.pdf

    work page 2023

  3. [3]

    Robinson, Byron Reeves, Nilam Ram, and Nick Haber

    Merve Cerit, Eric Zelikman, Mu-Jung Cho, Thomas N. Robinson, Byron Reeves, Nilam Ram, and Nick Haber. 2025. Media Content Atlas: A Pipeline to Explore and Investigate Multidimensional Media Space using Multimodal LLMs. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA ’25). Association for Computin...

    arXiv 2025

  4. [4]

    Communication and Marketing Staff. 2019. Improvement toward inclusion in film, but more work to be done. https://annenberg.usc.edu/news/research-and- impact/improvement-toward-inclusion-film-more-work-be-done Accessed: 16 October 2024

    work page 2019

  5. [5]

    Peter Zhuowei Gan, Arcot Sowmya, and Gelareh Mohammadi. 2023. CLIP-based Model for Effective and Explainable Apparent Personality Perception. In Proceed- ings of the 1st International Workshop on Multimodal and Responsible Affective Computing (Ottawa ON, Canada) (MRAC ’23). Association for Computing Ma- chinery, New York, NY, USA, 29–37. https://doi.org/1...

    work page arXiv 2023

  6. [6]

    GDI. 2021. Gender Bias and Inclusion in Advertising in India. https://geenadavisinstitute.org/research/gender-bias-and-inclusion-in- advertising-in-india/ p.7,17

    work page 2021

  7. [7]

    GDI. 2021. Women Over 50: The Right To Be Seen on Screen. https://geenadavisinstitute.org/research/women-over-50-the-right-to-be- seen-on-screen/

    work page 2021

  8. [8]

    GDI, UNICEF China, and UN Women China. 2024. Content Analysis of Gender Representation in Advertising in China: A Report Commissioned by UNICEF China and UN Women China. https://geenadavisinstitute.org/research/content- analysis-of-gender-representation-in-advertising-in-china/ p.10

    work page 2024

Show all 33 references
  1. [9]

    GDI and Lyda Hill Foundation. 2018. Portray her:Representations of Women STEM Characters in Media. https://geenadavisinstitute.org/research/portray- her/ Accessed: 10 April 2025

    2018

  2. [10]

    Google. 2014. The women missing from the silver screen and the technology used to find them. https://about.google/main/gender-equality-films/ Accessed: 15 January 2025

    2014

  3. [11]

    Narayanan

    Tanaya Guha, Che-Wei Huang, Naveen Kumar, Yan Zhu, and Shrikanth S. Narayanan. 2015. Gender Representation in Cinematic Content: A Multimodal Ap- proach. In Proceedings of the 2015 ACM on International Conference on Multimodal Interaction (Seattle, Washington, USA) (ICMI ’15)....

    2015

  4. [12]

    Rajat Hebbar, Krishna Somandepalli, and Shrikanth S Narayanan. 2018. Improving Gender Identification in Movie Audio Using Cross-Domain Data.. In Interspeech. 282–286

    2018

  5. [13]

    Kimmo Kärkkäinen and Jungseock Joo. 2019. FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age. CoRR abs/1908.04913 (2019). arXiv:1908.04913 http://arxiv.org/abs/1908.04913

    2019 arXiv

  6. [14]

    Kizilcec

    René F. Kizilcec. 2016. How Much Information? Effects of Transparency on Trust in an Algorithmic Interface. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (San Jose, California, USA) (CHI ’16). Association for Computing Machinery, New York, NY,...

    2016

  7. [15]

    Prakhar Kulshreshtha and Tanaya Guha. 2018. An Online Algorithm for Con- strained Face Clustering in Videos. In 2018 25th IEEE International Conference on Image Processing (ICIP). 2670–2674. https://doi.org/10.1109/ICIP.2018.8451343

    2018

  8. [16]

    Maksim Kuprashevich, Grigorii Alekseenko, and Irina Tolstykh. 2024. Beyond Specialization: Assessing the Capabilities of MLLMs in Age and Gender Estima- tion. arXiv:2403.02302 [cs.CV] https://arxiv.org/abs/2403.02302

    2024 arXiv

  9. [17]

    Antoine Mazières, Telmo Menezes, and Camille Roth. 2021. Computational appraisal of gender representativeness in popular movies. Humanities and Social Sciences Communications 137, 8 (2021)

    2021

  10. [18]

    Andrea Papenmeier, Gwenn Englebienne, and Christin Seifert. 2019. How model accuracy and explanation fidelity influence user trust. CoRR abs/1907.12652 (2019). arXiv:1907.12652 http://arxiv.org/abs/1907.12652

    2019 arXiv

  11. [19]

    Andrea Papenmeier, Dagmar Kern, Gwenn Englebienne, and Christin Seifert

  12. [20]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. CoRR abs/2103....

    2021 arXiv

  13. [21]

    Anil Ramakrishna, Victor R Martínez, Nikolaos Malandrakis, Karan Singla, and Shrikanth Narayanan. 2017. Linguistic analysis of differences in portrayal of movie characters. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Lo...

    2017

  14. [22]

    Rasmus Rothe, Radu Timofte, and Luc Van Gool. 2018. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision 126, 2-4 (2018), 144–157

    2018

  15. [23]

    Sefik Ilkin Serengil and Alper Ozpinar. 2020. LightFace: A Hybrid Deep Face Recognition Framework. In 2020 Innovations in Intelligent Systems and Applica- tions Conference (ASYU). IEEE, 23–27. https://doi.org/10.1109/ASYU50717.2020. 9259802

    2020

  16. [24]

    Rahul Sharma, Krishna Somandepalli, and Shrikanth Narayanan. 2019. Toward visual voice activity detection for unconstrained videos. In2019 IEEE International Conference on Image Processing (ICIP) . IEEE, 2991–2995

    2019

  17. [25]

    Smith, Marc Choueiti, Katherine Pieper, Kevin Yao, Ariana Case, and Angel Choi

    Stacy L. Smith, Marc Choueiti, Katherine Pieper, Kevin Yao, Ariana Case, and Angel Choi. 2019. Inequality in 1,200 Popular Films: Examining Portrayals of Gender, Race/Ethnicity, LGBTQ & Disability from 2007 to 2018. https: //assets.uscannenberg.org/docs/aii-inequality-report-2...

    2019

  18. [26]

    Martinez, Naveen Kumar, Hartwig Adam, and Shrikanth Narayanan

    Krishna Somandepalli, Tanaya Guha, Victor R. Martinez, Naveen Kumar, Hartwig Adam, and Shrikanth Narayanan. 2021. Computational Media Intelligence: Human-Centered Machine Analysis of Media. Proc. IEEE 109, 5 (2021), 891–

    2021

  19. [27]

    Swift and Ben Steeden

    Hannah J. Swift and Ben Steeden. 2020. Exploring representations of old age and ageing. https://ageing-better.org.uk/sites/default/files/2020-03/Exploring- representations-of-old-age.pdf Centre for Ageing Better

    2020

  20. [28]

    Williamson

    Evdoxia Taka, Sebastian Stein, and John H. Williamson. 2024. Does Interactive Conditioning Help Users Better Understand the Structure of Probabilistic Models? IEEE Transactions on Visualization and Computer Graphics 30, 7 (2024), 3256–3267. https://doi.org/10.1109/TVCG.2022.3231967

    2024

  21. [29]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. 2019. HuggingFace’s Transformers: State-of-the-art Natural Language Processing. CoRR abs/1910.03771 (2019). arXiv:1...

    2019 arXiv

  22. [30]

    Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, and Stan Z. Li

  23. [910]

    https://doi.org/10.1109/JPROC.2020.3047978

    2020

  24. [2017]

    In 2017 IEEE International Conference on Computer Vision (ICCV)

    S ˆ3FD: Single Shot Scale-Invariant Face Detector. In 2017 IEEE International Conference on Computer Vision (ICCV) . 192–201. https://doi.org/10.1109/ICCV. 2017.30

    2017 doi

  25. [2022]

    ACM Trans

    It’s Complicated: The Relationship between User Trust, Model Accuracy and Explanations in AI. ACM Trans. Comput.-Hum. Interact. 29, 4, Article 35 (March 2022), 33 pages. https://doi.org/10.1145/3495013

    2022 doi

Pith tools

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