Pith. sign in

REVIEW 4 major objections 5 minor 26 references

A Multi-Model Metric-based Selection Framework for Abstractive Text summarization

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Selecting the best of three candidate summaries by the average of ROUGE-L, BLEU, and BERTScore measured against the source article outperforms each constituent summarization model on CNN/DailyMail.

desk verdict The paper's own tables contradict its central claim: the base-line average is below T5-small, and the 'highest BERTScore' claim ignores DAMB in the same table. read the letter →

arxiv 2606.05494 v4 pith:UPYHCTT2 submitted 2026-06-03 cs.CL cs.AI

Ahmed Alansary , Ali Hamdi This is my paper
classification cs.CLcs.AI
keywords abstractivesummarizationmulti-modelensemblemetric-basedselectionROUGE-LBLEUBERTScoreCNN/DailyMailtransformerfine-tuning
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 tries to establish that a simple post-hoc selection rule can make multi-model summarization more robust than any single model: generate one candidate summary from each of three fine-tuned transformers (T5-small, PEGASUS-xsum, LED-base), score each candidate against the original article with the average of ROUGE-L, BLEU, and BERTScore, and output the highest-scoring candidate. On CNN/DailyMail, this Multi-Model Adaptive Summarization Framework (MASF) reports a fine-tuned BERTScore of 88.63% and higher ROUGE-L and BLEU than its individual models, and it is competitive with much larger LLMs. The authors argue that the framework leverages complementary strengths of different architectures and improves consistency across diverse articles. If true, the result matters because it suggests quality gains can come from selection rather than further model scale or training.

What carries the argument

The load-bearing mechanism is the selection rule defined by Eqs. (6)–(7): for each candidate summary S_i, compute Score(S_i) = (ROUGE-L(S_i) + BLEU(S_i) + BERTScore-F1(S_i)) / 3, where all three metrics are computed between the candidate and the source article, then choose S* = argmax Score(S_i). This composite score is meant to capture both lexical overlap and semantic similarity, and the argmax turns three candidate summaries into one final output. The framework also uses LoRA-based parameter-efficient fine-tuning for the LED-base model, but the selection rule is the element that carries the paper's argument.

What would settle it

For a held-out set of CNN/DailyMail test articles, compute the three candidate summaries, apply the MASF selection rule, and evaluate the selected summary with reference-based metrics (e.g., ROUGE-1/2/L and BERTScore against the human highlights). If the selected summaries do not beat the best single model's summaries on those reference-based metrics — or if a random/majority baseline does equally well — the central claim fails. Also check Table II's DAMB BERTScore of 88.70%; if it holds, the paper's 'highest BERTScore' claim is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that candidate summaries generated independently by three fine-tuned transformer models can be reliably ranked by a composite score — the simple average of ROUGE-L, BLEU, and BERTScore computed between each candidate and the source article — and that the argmax selection (Eq. 7) identifies the most informative summary. The authors report that in the fine-tuned setting MASF achieves 88.63% BERTScore, 32.75 ROUGE-L, and 16.00 BLEU, improving the average score over the strongest individual model, T5-small, by 3.01 percentage points, and state that it achieves the highest BERTScore among all compared methods, outperforming several LLMs such as GPT3-D2, Falcon-7b, and

Load-bearing premise

The framework's selection rule assumes that a candidate summary's average ROUGE-L, BLEU, and BERTScore against the source article is a reliable proxy for its quality relative to human-written reference summaries — an assumption the paper does not test.

Editorial extensions

If this is right

  • Combining a handful of small fine-tuned models plus a metric-based selector can match or beat much larger LLM summarizers on CNN/DailyMail, suggesting scale is not the only path to quality.
  • The selection rule works without human reference summaries at inference time, since it scores candidates against the source article, so it can be dropped into existing single-model pipelines at negligible cost.
  • Because the method is model-agnostic, it applies to any set of candidate summaries — from different prompts, decodings, or architectures — as a post-hoc robustness layer.
  • The consistent BERTScore gains imply the selector tends to prefer summaries semantically aligned with the source, which may translate to fewer hallucinations (though the paper does not test this directly).
  • The framework's competitive average score (45.80%) against BRIO (46.00%) with smaller models suggests lightweight ensembles could be a practical alternative to training larger summarizers.

Reading between the lines

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

  • The selection metric is article-relative, not reference-relative; the paper gives no evidence that scoring high against the source correlates with scoring high against human highlights, so the framework may in practice select the most extractive or copy-heavy candidate. A direct test would compare selected summaries to human references across the test set.
  • The paper's 'highest BERTScore' statement is, on its face, inconsistent with its own Table II, which lists DAMB at 88.70% versus MASF's 88.63%; if that table is accurate, the claim should be read as 'highest among the models, excluding DAMB' or the comparison is in error.
  • A natural extension, mentioned in the paper's future work, is to swap the article-relative metrics for reference-free, human-aligned or factuality metrics; this would test whether the selection mechanism generalises beyond lexical overlap to true summary quality.
  • The largest fine-tuned gain appears in BLEU (+3.5 points over T5-small), a lexical-precision metric measured against the article; this is consistent with the extractive-bias concern and suggests the selector may disproportionately favour candidates that reuse source phrases.
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

4 major / 5 minor

Summary. The paper proposes a Multi-Model Adaptive Summarization Framework (MASF) that generates three candidate summaries from fine-tuned T5-small, PEGASUS-xsum, and LED-base models, scores each candidate with an average of ROUGE-L, BLEU, and BERTScore, and selects the highest-scoring summary as the final output. The models are fine-tuned on CNN/DailyMail and compared against the individual models and several prior/LLM baselines. The central claims are that MASF achieves the highest BERTScore among all compared methods (88.63%) and consistently outperforms the individual baseline and fine-tuned summarization models in overall summarization quality.

Significance. The idea of combining multiple summarization models with a lightweight metric-based selector is practically appealing and could be a useful robustness technique if it works. The paper uses a standard public dataset, standard fine-tuned models, and a simple transparent selection rule, which are strengths in terms of reproducibility. However, the reported results do not support the stated claims: the paper's own tables contradict the claimed BERTScore superiority and the claimed consistent outperformance over baselines. Moreover, because the selection metric is computed against the source article rather than human reference summaries, and the same metric is then used to report system performance, the reported gains are partly an artifact of the selection operator. These issues are load-bearing and make the current manuscript unsuitable for publication in its present form.

major comments (4)
  1. [Table III / Section V] The central claim that MASF 'consistently outperforms the individual baseline and fine-tuned summarization models' is contradicted by the paper's own Table III in the baseline setting. Using the paper's average of ROUGE-L, BLEU, and BERTScore as in Eq. (6), T5-small receives (86.31 + 25.19 + 7.75)/3 = 39.75%, while MASF receives (87.07 + 24.30 + 7.40)/3 = 39.59%. Section V explicitly acknowledges that T5-small has higher ROUGE-L and BLEU, yet the Conclusion asserts consistent outperformance. This internal inconsistency directly undermines the main contribution.
  2. [Abstract / Section V / Table II] The abstract and Section V state that MASF 'achieves the highest BERTScore among all compared methods with a score of 88.63%.' Table II lists DAMB with a BERTScore of 88.70%, which is higher than 88.63%. Under the paper's own comparison table, the claimed 'highest' BERTScore is false. This is not a cosmetic wording issue: it pertains to the primary reported result.
  3. [Eq. (6), Eq. (7), Tables III–IV] The selection rule in Eq. (7) chooses the candidate with the highest average of ROUGE-L, BLEU, and BERTScore, where Eq. (6) states these scores are computed between the generated summary and the article, not against reference summaries. The same article-relative scores are then reported in Tables III–IV as MASF's performance. This creates a circularity: the selected summary's score is, by construction, at least as high as each candidate's score at the instance level, so reporting these scores as evidence of improved summary quality is partly a mathematical consequence of the argmax. Furthermore, in the baseline table, MASF's average (39.59%) is below T5-small's average (39.75%); if the reported averages were article-relative and selection used Eq. (7), the selected average would have to be at least as large as every candidate's average. The observed violation suggests either the table sc
  4. [Section IV-D / Selection metric validity] Even setting aside the internal contradictions, the framework's validity depends on whether article-relative ROUGE-L/BLEU/BERTScore is a proxy for reference-relative summary quality. The paper provides no evidence, such as correlation with reference-based metrics or human judgments, that selecting the candidate most similar to the source article yields summaries of higher quality. Without such evidence, the method may simply favor extractive or source-copying candidates, and the reported 'gains' over fine-tuned models are not established.
minor comments (5)
  1. [Throughout] Use consistent terminology: the title, abstract, and body alternate among 'Metric-based Selection', 'Adaptive Selection', and 'Multi-Model Adaptive Summarization Framework'. This makes the contribution harder to identify.
  2. [Tables II–IV] The table columns do not specify whether scores are averages over the test set with standard deviations, and no significance tests are reported. Some values are missing (e.g., DAMB's BLEU in Table II), and the 'AverageScore' column is not defined as the mean of the three preceding columns.
  3. [Section V] The phrase 'higher than all reported models in the comparison' is immediately followed by the observation that text-davinci-003 has a higher average score (53.75%), which is confusing. Clarify whether BERTScore or average score is the claimed ranking criterion.
  4. [Table I] There are typos in the article sample text, e.g., 'Firoentina' and 'PSG and Spanish clubs' with inconsistent capitalization. These do not affect the technical content but should be cleaned.
  5. [References] Several references are incompletely formatted (e.g., [19], [26] lack page numbers or venue details in places), and the reference list numbering should be checked against in-text citations.

Circularity Check

0 steps flagged · score 0.0 of 10

No construction-level circularity; selection score and reported evaluation are not the same quantity.

full rationale

The only candidate circularity is the relationship between Eq. (6)'s selection score and the reported ROUGE-L/BLEU/BERTScore results. The paper defines the selection score explicitly as article-relative: 'ROUGE-L measures lexical overlap between the generated summary and the article...' (Section IV-D), so Eq. (7) maximizes a source-article score, not a reference-based evaluation score. The performance tables compare against published CNN/DailyMail results, which are reference-based, and the paper never states that Table III/IV values are the same Eq. (6) quantities. This is confirmed by Table III itself: MASF's average is 39.59%, below T5-small's 39.75%; if the reported table were simply the average of the selection score being maximized, the argmax operator would force MASF's average to be at least every candidate's average. Hence no equation-level reduction exists. No parameter is fitted to the reported test metrics; the selection rule is fixed. The self-citations in the reference list (e.g., [11], [13], [19]) appear only in related work and do not justify the central MASF derivation. The abstract's 'highest BERTScore... 88.63%' appears contradicted by DAMB's 88.70% in Table II, and Table III conflicts with the claim that MASF outperforms all base-line models; those are internal-consistency/empirical problems, not circularity. Overall, the derivation is not circular; it may be empirically unreliable, but that is a correctness risk outside this pass.

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

The central claim rests on an unvalidated proxy (article-relative metrics), an assumption of comparability with external results, and a diversity assumption; no new entities are introduced.

free parameters (3)
  • Selection metric weights (ROUGE-L, BLEU, BERTScore) = 1/3, 1/3, 1/3 (equal-weight average, Eq. 6)
    The composite selection score is a simple average of three metrics chosen without ablation; the central claim that selection improves quality depends on this weighting.
  • Maximum generation length = 128 tokens
    Fixed for all models; affects summary length and scores, not varied or justified.
  • LoRA fine-tuning hyperparameters for LED-base = not reported
    LoRA rank and related settings are not given; needed to reproduce the fine-tuned LED-base and thus the framework.
assumptions (3)
  • domain assumption ROUGE-L, BLEU and BERTScore computed between a candidate summary and the source article (Eq. 6) are valid proxies for summary quality against human reference summaries.
    Used in §IV-D/E as the selection criterion; if false, the argmax in Eq. 7 may select copy-heavy summaries that score well against the article but poorly against references.
  • domain assumption Results from cited works in Table II are directly comparable to MASF (same CNN/DailyMail split, same metric implementations).
    Table II mixes scores from different papers without confirming identical evaluation protocols; if protocols differ, the comparison is invalid.
  • domain assumption The three fine-tuned models produce sufficiently diverse candidate summaries for per-article selection to improve aggregate quality.
    No diversity analysis is provided; if candidates are near-duplicates, selection cannot explain the claimed gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Multi-Model Metric-based Selection Framework for Abstractive Text summarization." pith.science (2026). https://pith.science/paper/UPYHCTT2

@misc{pith2026260605494,
  author       = {Pith},
  title        = {Pith review of: A Multi-Model Metric-based Selection Framework for Abstractive Text summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UPYHCTT2}},
  note         = {Machine review of arXiv:2606.05494}
}
read the original abstract

Automatic text summarization has become increasingly important due to the rapid growth of digital textual information. This paper presents a Multi-Model Summarization Framework designed to improve the robustness and quality of abstractive text summarization. Relying on a single model often leads to inconsistent summarization quality across articles with varying structures and topics. To address this limitation, the proposed framework integrates multiple fine-tuned transformer-based summarization models and introduces a metric-based selection mechanism. In this framework, each model independently generates a candidate summary for the same input article. The generated summaries are then evaluated using automatic evaluation metrics that capture both lexical similarity and semantic relevance. Based on these scores, the framework selects the highest-quality summary as the final output. The models are fine-tuned and evaluated on the widely used CNN/DailyMail news summarization dataset. Experimental results demonstrate that the proposed framework achieves the highest BERTScore among all compared methods with a score of 88.63%. It also outperforms several LLMs such as GPT3-D2, Falcon-7b, and Mpt-7b, highlighting its effectiveness and robustness. These findings highlight the effectiveness of leveraging multiple transformer-based models within a metric-based selection strategy to improve the quality and robustness of automatic text summarization systems.

Figures

Figures reproduced from arXiv: 2606.05494 by the authors.

Figure 1
Figure 1. Overview of the proposed multi-model adaptive sum [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Overview of the proposed multi-model adaptive summarization [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Visual Comparison between baseline models and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (2 more)
Figure 2
Figure 2. Figure 2: Visual Comparison between base-line models and MASF [PITH_FULL_IMAGE:figures/full_fig_p004_2.png]
Figure 3
Figure 3. Figure 3: Visual Comparison between fine-tuned models and MASF [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

Review history (4 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 2 linked inside Pith

  1. [1]

    Summn: A multi-stage summarization frame- work for long input dialogues and documents,

    Y . Zhang, A. Ni, Z. Mao, C. H. Wu, C. Zhu, B. Deb, A. Awadallah, D. Radev, and R. Zhang, “Summn: A multi-stage summarization frame- work for long input dialogues and documents,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 1592–1604

    2022

  2. [2]

    Chatgpt vs human-authored text: Insights into controllable text summarization and sentence style transfer,

    D. Liu and V . Demberg, “Chatgpt vs human-authored text: Insights into controllable text summarization and sentence style transfer,” inProceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), 2023, pp. 1–18

    2023

  3. [3]

    Domain adaptation with pre-trained transformers for query-focused abstractive text summariza- tion,

    M. T. R. Laskar, E. Hoque, and J. X. Huang, “Domain adaptation with pre-trained transformers for query-focused abstractive text summariza- tion,”Computational Linguistics, vol. 48, no. 2, pp. 279–320, 2022

    2022

  4. [4]

    Text summarization using large language models: A comparative study of mpt-7b-instruct, falcon- 7b-instruct, and openai chat-gpt models,

    L. Basyal and M. Sanghvi, “Text summarization using large language models: A comparative study of mpt-7b-instruct, falcon- 7b-instruct, and openai chat-gpt models,” 2023. [Online]. Available: https://arxiv.org/abs/2310.10449

    arXiv 2023

  5. [5]

    Abstractive meeting summarization: A survey,

    V . Rennard, G. Shang, J. Hunter, and M. Vazirgiannis, “Abstractive meeting summarization: A survey,”Transactions of the Association for Computational Linguistics, vol. 11, pp. 861–884, 07 2023

    2023

  6. [6]

    Anlirika: An LSTM–CNN flow twister for spoken language identification,

    A. Scherbakov, L. Whittle, R. Kumar, S. Singh, M. Coleman, and E. Vy- lomova, “Anlirika: An LSTM–CNN flow twister for spoken language identification,” inProceedings of the Third Workshop on Computational Typology and Multilingual NLP, E. Vylomova, E. Salesky, S. Mielke, G. Lapesa, R. Kumar, H. Hammarstr ¨om, I. Vuli ´c, A. Korhonen, R. Re- ichart, E. M. P...

    2021

  7. [7]

    Riro: Reshaping inputs, refining outputs unlocking the potential of large language models in data-scarce contexts,

    A. Hamdi, H. Kassab, M. Bahaa, and M. Mohamed, “Riro: Reshaping inputs, refining outputs unlocking the potential of large language models in data-scarce contexts,” inThe International Conference of Advanced Computing and Informatics. Springer Nature Switzerland Cham, 2024, pp. 69–79

    2024

  8. [8]

    Au- tomatic text summarization: A comprehensive survey,

    W. S. El-Kassas, C. R. Salama, A. A. Rafea, and H. K. Mohamed, “Au- tomatic text summarization: A comprehensive survey,”Expert systems with applications, vol. 165, p. 113679, 2021

    2021

Show all 26 references
  1. [9]

    A comprehensive survey on automatic text summarization with exploration of llm-based methods,

    Y . Zhang, H. Jin, D. Meng, J. Wang, and J. Tan, “A comprehensive survey on automatic text summarization with exploration of llm-based methods,”Neurocomputing, p. 131928, 2025

    2025

  2. [10]

    A survey of automatic text summarization using graph neural networks,

    M. F. Salchner and A. Jatowt, “A survey of automatic text summarization using graph neural networks,” inProceedings of the 29th International Conference on Computational Linguistics, N. Calzolari, C.-R. Huang, H. Kim, J. Pustejovsky, L. Wanner, K.-S. Choi, P.-M. Ryu, H.-H. Che...

    2022

  3. [11]

    Balancing factual con- sistency and diversity in abstractive summarization via model-agnostic composite reranking,

    M. Elewa, A. Hamdi, H. Kassab, and K. Shaban, “Balancing factual con- sistency and diversity in abstractive summarization via model-agnostic composite reranking,” in2025 IEEE/ACS 22nd International Conference on Computer Systems and Applications (AICCSA). IEEE, 2025, pp. 1– 8

    2025

  4. [12]

    Abstractive text summarization: State of the art, challenges, and improvements,

    H. Shakil, A. Farooq, and J. Kalita, “Abstractive text summarization: State of the art, challenges, and improvements,”Neurocomputing, vol. 603, p. 128255, 2024

    2024

  5. [13]

    Lexisem: A re- ranker balancing lexical and semantic quality for enhanced abstractive summarization,

    E. Aloraini, H. Kassab, A. Hamdi, and K. Shaban, “Lexisem: A re- ranker balancing lexical and semantic quality for enhanced abstractive summarization,”Neurocomputing, vol. 650, p. 130816, 2025

    2025

  6. [14]

    Automatic text summarization using soft- cosine similarity and centrality measures,

    M. Jain and H. Rastogi, “Automatic text summarization using soft- cosine similarity and centrality measures,” in2020 4th International Conference on Electronics, Communication and Aerospace Technology (ICECA), 2020, pp. 1021–1028

    2020

  7. [15]

    Ranksum—an unsupervised extractive text summarization based on rank fusion,

    A. Joshi, E. Fidalgo, E. Alegre, and R. Alaiz-Rodriguez, “Ranksum—an unsupervised extractive text summarization based on rank fusion,” Expert Systems with Applications, vol. 200, p. 116846, 2022

    2022

  8. [16]

    Extractive summarization using extended TextRank algorithm,

    A. N. V ora, R. M. Jain, A. S. Shah, and S. Sonawane, “Extractive summarization using extended TextRank algorithm,” inProceedings of the 21st International Conference on Natural Language Processing (ICON), S. Lalitha Devi and K. Arora, Eds. AU-KBC Research Centre, Chennai, Ind...

    2024

  9. [17]

    A topic modeled un- supervised approach to single document extractive text summarization,

    R. Srivastava, P. Singh, K. Rana, and V . Kumar, “A topic modeled un- supervised approach to single document extractive text summarization,” Knowledge-Based Systems, vol. 246, p. 108636, 2022

    2022

  10. [18]

    Bidirectional lstm networks for abstractive text summarization,

    A. Kova ˇcevi´c and D. Ke ˇco, “Bidirectional lstm networks for abstractive text summarization,” inInternational Symposium on Innovative and Interdisciplinary Applications of Advanced Technologies. Springer, 2021, pp. 281–293

    2021

  11. [19]

    Summfactscore: A claim-centric framework forreference-free factual consistency evalu- ation inlong-document summarization,

    E. Aloraini, A. Hamdi, E. Elmahjubet al., “Summfactscore: A claim-centric framework forreference-free factual consistency evalu- ation inlong-document summarization,”Ali and Elmahjub, Ezieddin, SummFactScore: A Claim-Centric Framework forReference-Free Factual Consistency Eval...

  12. [20]

    Discourse-aware neural extractive text summarization,

    J. Xu, Z. Gan, Y . Cheng, and J. Liu, “Discourse-aware neural extractive text summarization,” inProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, Eds. Online: Association for Computatio...

    2020

  13. [21]

    Deep learning for text summarization using nlp for automated news digest,

    K. Rani Krishna, K. Somasundaram, P. Arulmozhivarman, S. A. Im- manuel, and E. Rajkumar, “Deep learning for text summarization using nlp for automated news digest,”Scientific Reports, vol. 15, no. 1, p. 36343, 2025

    2025

  14. [22]

    Extractive summarization as text matching,

    M. Zhong, P. Liu, Y . Chen, D. Wang, X. Qiu, and X. Huang, “Extractive summarization as text matching,” inProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, Eds. Online: Association for ...

    2020

  15. [23]

    Adapted large language models can outperform medical experts in clinical text summarization,

    D. Van Veen, C. Van Uden, L. Blankemeier, J.-B. Delbrouck, A. Aali, C. Bluethgen, A. Pareek, M. Polacin, E. P. Reis, A. Seehofnerov ´aet al., “Adapted large language models can outperform medical experts in clinical text summarization,”Nature medicine, vol. 30, no. 4, pp. 1134...

    2024

  16. [24]

    A survey on cross-lingual summarization,

    J. Wang, F. Meng, D. Zheng, Y . Liang, Z. Li, J. Qu, and J. Zhou, “A survey on cross-lingual summarization,”Transactions of the Association for Computational Linguistics, vol. 10, pp. 1304–1323, 2022

    2022

  17. [25]

    News summarization and evaluation in the era of gpt-3,

    T. Goyal, J. J. Li, and G. Durrett, “News summarization and evaluation in the era of gpt-3,” 2023. [Online]. Available: https://arxiv.org/abs/2209.12356

    2023 arXiv

  18. [26]

    Damb: A dynamic adaptive multi-model benchmarking framework for abstractive text summarization,

    M. Burukanli and D. Ari, “Damb: A dynamic adaptive multi-model benchmarking framework for abstractive text summarization,” 12 2025

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.