Pith. sign in

REVIEW 4 major objections 5 minor 31 references

MIDAS: Multi-LLM Iterative Data-Adaptive Summarization

T0 review · 4 major / 5 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 MIDAS claims that grounding prompt optimization in reference-derived formatting rules, rather than generic critique alone, yields the strongest enterprise summarization prompts across five output formats.

desk verdict A solid incremental prompt-optimization method whose reported gains outrun the statistics: the test set is too small and the iteration count was tuned on it. read the letter →

arxiv 2608.04307 v1 pith:CE527YHF submitted 2026-08-05 cs.CL cs.AIcs.MA

Karen Lee , Dhanashree Balaram , Seojun Shon , Umair Rasheed This is my paper · ORCID
classification cs.CLcs.AIcs.MA
keywords LLMsummarizationpromptoptimizationenterprisesupportticketspolicylearningcritique-drivenrefinementmulti-LLMROUGEBERTScore
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

MIDAS is an automatic prompt optimizer for enterprise text summarization: instead of refining prompts with generic critique criteria, it first induces a 'policy block' of formatting rules from a held-out sample of reference summaries, then iterates among a generator, a scorer, and a single chain-of-thought critic/prompt-rewriter. The paper claims this data-grounded loop is what lets prompts adapt to diverse output schemas—structured metadata, subject lines, multilingual free text, English-normalized text, and tag lists—without manual tuning. Applied to about 24,600 IT help-desk tickets, MIDAS reports the strongest overall ROUGE-1/2/L and BERTScore F1 across five formats, with gains over CriSPO up to 11.0/18.2/8.0 percent, and shows the same advantage on a finance earnings-call benchmark and in a heterogeneous multi-LLM configuration. A sympathetic reader would take the central claim to be: reference-derived structural patterns, not longer or more elaborate critique, are what drive prompt optimization gains in summarization.

What carries the argument

The load-bearing mechanism is the policy block: a set of natural-language rules induced by a dedicated Data Pattern LLM from a stratified holdout subset of reference summaries, then injected into the critic template so every refinement step is grounded in the target domain's observed formatting conventions (e.g., 'Type must be exactly one of Incident, Request, Change, Problem'). Around it sits a three-agent loop—generator, scorer LLM with configurable dimension weights, and a unified chain-of-thought critic that combines critique and suggestion into one LLM call—which eliminates the separate suggestion stage of CriSPO and reduces inference overhead. The policy block is what turns generic prompt optimization into data-adaptive prompt optimization.

What would settle it

Evaluate MIDAS against CriSPO and ZERA on a summarization benchmark where every reference summary is written independently by human annotators (no LLM-generated references), using the same 30-iteration budget; if the ROUGE and BERTScore advantages shrink to statistical noise, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that critique-driven prompt optimization is bottlenecked by task-agnostic critique dimensions, and that extracting dataset-specific structural rules from reference summaries—encoded as a policy block—removes the bottleneck. MIDAS operationalizes this with three coordinated agents: a generator, an LLM scorer that returns weighted dimension scores, and an implicit chain-of-thought critic that emits both critique and a full revised prompt in a single pass, given the policy block, generated examples, and optimization history. The paper reports that this architecture outperforms zero-shot, in-context learning, CriSPO (including a 100-iteration control), and ZERA across four of five output formats, ties on the fifth while taking the best BERTScore, and retains most of its edge with mixed foundation models and on ECTSum. The strongest single example is Format 5 (structured tag output), where MIDAS lifts ROUGE-1 from 74.6 to 82.8 and ROUGE-2 from 54.1 to 63.9 relative to CriSPO.

Load-bearing premise

The reference summaries for the two free-text formats are generated by GPT-5.2 from ticket tags and then human-verified, and the generator, critic, scorer, and data-pattern LLM are all GPT-5.2, so the reported free-text gains may partly reflect the optimizer aligning outputs with the same model family's own distribution rather than with independent human quality.

Editorial extensions

If this is right

  • MIDAS's prompt optimization works with as few as 40 training examples and 200 stratified pattern samples, so organizations with small labeled ticket sets can still get format-adapting prompts.
  • Merging critique and prompt rewriting into one LLM call halves the refinement pipeline relative to CriSPO, cutting inference overhead per iteration.
  • The 100-iteration CriSPO control performing worse than its 30-iteration run implies that compute alone does not buy quality; data-grounded constraints are what matter.
  • Heterogeneous multi-LLM assignments (different models for scoring, critique, and generation) preserve most of the gain, suggesting the method is not tied to one model family.
  • Gains transfer to a finance-domain benchmark (ECTSum), indicating the policy-block mechanism generalizes beyond IT tickets.

Reading between the lines

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

  • If the policy block is the active ingredient, then a controlled ablation that replaces it with random formatting rules should destroy the gains; the paper does not report this ablation, so the claim rests on the architecture comparison.
  • Because Formats 3 and 4 are evaluated against GPT-5.2-generated references, the measured free-text gains may partly measure self-consistency with the generator's distribution; an independent human-written reference set would separate structural learning from style alignment.
  • The same policy-learning loop could be applied to other structured generation tasks (e.g., legal clause extraction, incident report formatting) where reference corpora already exist, without task-specific prompt design.
  • The distribution-aware sampling used for pattern extraction could be tested for sensitivity: whether imbalance in the holdout subset changes which rules are induced and how fast the optimization converges.
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 MIDAS, a multi-LLM iterative prompt-optimization framework for enterprise summarization. MIDAS learns a policy block of domain-specific formatting rules from reference summaries, then iteratively refines a task prompt through a unified chain-of-thought critic and scorer. The method is evaluated on a multilingual IT help desk ticket dataset across five output formats (structured metadata, subject lines, multilingual free text, English-normalized free text, and keyword tags), against Zero-Shot, ICL, CriSPO, CriSPO with 100 iterations, and ZERA baselines. The authors report consistent gains over baselines for most formats, with ROUGE-1/2/L improvements over CriSPO of up to 11.0%, 18.2%, and 8.0%, and they additionally report a cross-domain experiment on the ECTSum finance benchmark. The paper includes prompt templates in the appendix and describes the data splits and optimization procedure.

Significance. If the central claim holds, MIDAS is a practically relevant contribution: it extends critique-driven prompt optimization with data-driven policy learning, consolidates critique and prompt refinement into a single LLM call, and demonstrates some robustness across output formats, model assignments, and domains. The paper is clearly written, the proposed architecture is well specified, the prompt templates are included, and the evaluation spans five output formats plus a second domain. However, the empirical evidence as presented is not yet sufficient to establish the central claim of consistent, statistically reliable gains. The main weaknesses are the small test set, the use of the test set for selecting the iteration count, the absence of variance or confidence intervals, and the unreported large-scale holdout validation. These issues are fixable within the manuscript's scope, but they affect every headline number.

major comments (4)
  1. [§5.2 and Figure 2] The optimization iteration count N=30 is selected via ablation on the same 30-sample test set used for the headline results in Table 2. Since the test set is used to choose the hyperparameter, the reported scores are optimistically biased and the comparison to baselines is not a fair evaluation. Please select N on a separate validation set (or use nested cross-validation) and report results on a test set that is untouched by any selection decision.
  2. [§5.2 and Table 2] All headline results are computed on only 30 test samples with two random seeds, and no per-seed breakdowns, bootstrap intervals, confidence intervals, or significance tests are reported. Many of the claimed margins are small (for example, Format 2 ROUGE-2: MIDAS 0.1699 vs. ICL 0.1670), and without uncertainty estimates these differences cannot be distinguished from sampling noise. Please report variance-aware statistics, at minimum per-seed results and bootstrap confidence intervals, and preferably significance tests.
  3. [§6, Large-Scale Validation] The paper states that the best-performing prompts were validated on the full 24,355-ticket holdout corpus and that results were consistent with optimization-time findings, but no numbers, tables, or intervals are provided for this validation. This is the only evidence that the gains generalize beyond the 30-sample test set, and the abstract and conclusion rely on it for the generalization claim. Please report the full holdout results, including the comparison against baselines on this large set.
  4. [§5.1, Table 1 note] For Format 3 and Format 4, the reference summaries are generated by GPT-5.2 from ticket tags and then human-verified, while the generator, critic, scorer, and data-pattern learner in the main MIDAS configuration all use GPT-5.2. The policy block is induced from these same reference summaries, so gains on the free-text formats may partly reflect self-consistency with the model family's own output conventions rather than improved external summarization quality. Please provide evidence that the human verification is reliable (for example, inter-annotator agreement) or supplement the free-text evaluation with an independently authored human reference set.
minor comments (5)
  1. [§5.2, first paragraph] The sentence 'All experiments use GPT-5.2 as the backbone model for summary generation, critique, scoring, and data pattern learning' is contradicted later by the multi-LLM configuration, which uses Claude models for some agents. Please clarify that this sentence describes the default single-backbone configuration.
  2. [§4.2] The description of the fast evaluation subset says 'default: 25 examples out of 40', but the experimental setup in §5.2 states a fixed pool of 40 training examples and 10 development examples. Please clarify the relationship between the training pool, development set, and fast-evaluation subset, and report the value actually used in the experiments.
  3. [Figure 2 caption] The caption states that metrics are computed on a 30-sample test set; since this is the same set used for hyperparameter selection, the caption should explicitly note that this is an ablation on the test set and that final results therefore require independent validation.
  4. [Table 2 and Figure 2] The table reports BERTScore F1 while Figure 2 lists BERTScore P/R/F1; please make the metric definitions consistent and specify which BERTScore variant (precision, recall, or F1) is reported in Table 2.
  5. [References] Reference 8 duplicates Reference 7 (both are Lin's ROUGE paper); please merge or disambiguate the two entries.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity on GPT-5.2-generated free-text reference formats; structured-data results remain independent.

  1. fitted input called prediction [Section 5.1, Table 1 note; Section 4.1 and 5.1 (GPT-5.2 backbone, Data Pattern LLM)]
    "Format Type3 and Format Type4 reference summaries are generated by GPT-5.2 conditioned on the ticket’s tag fields (grounded generation), and subsequently verified by human annotators to ensure factual consistency with the source ticket. ... All experiments use GPT-5.2 as the backbone model for summary generation, critique, scoring, and data pattern learning."

    For Formats 3 and 4, the summaries used as ground truth in ROUGE/BERTScore evaluation are themselves produced by GPT-5.2, the same model family used as the summary generator, critic, scorer, and Data Pattern LLM. The Data Pattern LLM induces the policy block from samples of these GPT-5.2 reference summaries, and the optimization loop maximizes a GPT-5.2 LLM score comparing generated output to these references. The reported 'predictions' for these two formats therefore partly amount to matching the generator's own output conventions: the evaluation target and the optimizing model share the same underlying distribution. Human verification only checks factual consistency, not wording or style, so the lexical target remains the reference generator's own output.

full rationale

The MIDAS derivation chain is otherwise self-contained: the policy block is induced from a disjoint 200-sample holdout, the prompts are optimized against a fixed training pool, and the strongest comparisons against CriSPO and ZERA on structured formats (1, 2, 5) do not reduce to the model's own outputs. No load-bearing self-citations, imported uniqueness theorems, or ansatz-by-citation are present. The main circularity risk is limited to Formats 3 and 4, where the evaluation references are GPT-5.2-generated and all optimizing agents are GPT-5.2, so the reported free-text ROUGE gains partially measure self-consistency with the reference generator's output distribution. Separate statistical concerns (N=30 selected on the 30-sample test set, no confidence intervals, claimed 24,355-ticket validation without reported numbers) are validity issues rather than definitional circularity; they lower confidence in the headline gains but do not make the derivation equivalent to its inputs.

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

The central evaluation depends on synthetic reference summaries for two formats, an LLM-based scorer from the same model family, manually chosen weights, and a test-set-selected iteration count. These are assumptions and fitted choices rather than external constraints.

free parameters (4)
  • Scorer dimension weights (w_core, w_unsupported, w_format) = (0.4, 0.3, 0.3)
    Chosen by hand in Section 4.2 based on task-specific prioritization; they define the optimization objective though not the reported ROUGE metrics.
  • Optimization iteration count N = 30
    Selected via ablation over N in {10,20,30,50} evaluated on the 30-sample test set (Figure 2, Section 5.2), which risks test-set-driven hyperparameter selection.
  • Fast evaluation subset size = 25 examples out of 40 training examples
    Default stated in Section 4.2; affects candidate prompt ranking and therefore which prompt is selected.
  • Random seeds = 2
    Two random seeds are used with no variance reported, making the stability of the reported gains unclear.
assumptions (4)
  • domain assumption ROUGE and BERTScore against the provided references are valid proxies for enterprise summarization quality.
    The paper reports only automatic metrics and cites prior work for LLM evaluator alignment; no human evaluation is conducted. Enters in Section 6 results.
  • domain assumption Reference summaries for Formats 3 and 4, generated by GPT-5.2 from tags and human-verified, represent correct summaries.
    If these references reflect GPT-5.2 style rather than human quality, free-text gains are partially circular. See Section 5.1 and Table 1.
  • domain assumption A GPT-5.2 LLM-based evaluator aligns with human judgment.
    The Scorer LLM uses the same backbone and Section 4.2 relies on prior work [9] to justify LLM evaluators.
  • ad hoc to paper The five output formats chosen by the authors map to real enterprise requirements.
    The format configurations are manually defined in Section 5.1 and are not independently established as representative enterprise summarization tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MIDAS: Multi-LLM Iterative Data-Adaptive Summarization." pith.science (2026). https://pith.science/paper/CE527YHF

@misc{pith2026260804307,
  author       = {Pith},
  title        = {Pith review of: MIDAS: Multi-LLM Iterative Data-Adaptive Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CE527YHF}},
  note         = {Machine review of arXiv:2608.04307}
}
read the original abstract

Text summarization is deceptively difficult. While condensing information seems straightforward, real-world enterprise summarization of support tickets, legal documents, incident reports, and more, demands strict adherence to domain-specific guidelines, output formats, and organizational conventions. Crafting prompts that reliably satisfy these constraints is labor-intensive, requiring significant human expertise and continuous maintenance as requirements evolve. Existing automated prompt optimization methods reduce this burden through Large Language Model (LLM) critique-driven refinement, yet remain limited by static prompts that cannot adapt to the diversity of summary applications. We propose Multi-LLM Iterative Data-Adaptive Summarization (MIDAS), a multi-LLM framework that extends this paradigm with data-driven pattern learning and use-case-specific personalization, enabling automatic adaptation to different summarization requirements without manual prompt engineering. Applied to enterprise customer ticket summarization across five output formats, MIDAS achieves the strongest overall performance against state-of-the-art critique-driven optimization frameworks such as CriSPO and ZERA, improving ROUGE-1 by up to 11.0%, ROUGE-2 by up to 18.2%, and ROUGE-L by up to 8.0%, while consistently improving BERTScore F1 across all formats and output types. We additionally demonstrate cross-model and cross-domain generalization through multi-LLM configurations and finance-domain summarization benchmarks.

Figures

Figures reproduced from arXiv: 2608.04307 by the authors.

Figure 1
Figure 1. Overview of MIDAS, an iterative multi-agent framework for automated prompt refinement. An LLM Summary Generator produces summaries that are evaluated by a Scorer LLM. A Data Pattern LLM extracts dataset-specific structural rules from a stratified holdout subset. These signals, together with optimization history, are pro￾vided to an Implicit Chain-of-Thought Critic and Prompt LLM, which generates cri￾tiques and refin… view at source ↗
Figure 2
Figure 2. Effect of optimization iterations on summarization quality. ROUGE-1/2/L and BERTScore P/R/F1 across N ∈ {10, 20, 30, 50} iterations on a 30-sample test set. N=30 achieves the best performance across all metrics [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [1]

    Dataset, Kaggle (2026), https://www.kaggle.com/datasets/tobiasbueck/multilingual-customer-support- tickets

    Bück, T.: Multilingual customer support tickets. Dataset, Kaggle (2026), https://www.kaggle.com/datasets/tobiasbueck/multilingual-customer-support- tickets

    work page 2026

  2. [2]

    arXiv preprint arXiv:2510.13907 (2025)

    Chen, Y., et al.: PDO: Prompt-dueling optimization for label-free LLM evaluation. arXiv preprint arXiv:2510.13907 (2025)

    arXiv 2025

  3. [3]

    Technical report, Forrester Consulting (2023)

    Forrester Consulting, Zendesk: The total economic impact™of zendesk. Technical report, Forrester Consulting (2023)

    work page 2023

  4. [4]

    In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)

    Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., Yang, Y.: Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. In: Proceedings of the International Conference on Learning Representations (ICLR) (2024)

    work page 2024

  5. [5]

    In: AAAI Conference on Artificial Intelligence (2024), https://api.semanticscholar.org/CorpusID:273098751

    He, H., Liu, Q., Xu, L., Shivade, C.P., Zhang, Y., Srinivasan, S., Kirchhoff, K.: Crispo: Multi-aspect critique-suggestion-guided automatic prompt optimiza- tion for text generation. In: AAAI Conference on Artificial Intelligence (2024), https://api.semanticscholar.org/CorpusID:273098751

    work page 2024

  6. [6]

    In: Proceedings of the KDD Workshop on Generative AI for Recommendation and Personalization (GenAIRecP) (2024)

    Hebert, L., Sayana, K., Jash, A., Karatzoglou, A., Sodhi, S., Doddapaneni, S., Cai, Y., Kuzmin, D.: PERSOMA: Personalized soft prompt adapter architecture for personalized language prompting. In: Proceedings of the KDD Workshop on Generative AI for Recommendation and Personalization (GenAIRecP) (2024)

    work page 2024

  7. [7]

    In: Text Summarization Branches Out

    Lin, C.Y.: ROUGE: A package for automatic evaluation of summaries. In: Text Summarization Branches Out. pp. 74–81. Association for Computational Linguis- tics, Barcelona, Spain (2004), https://aclanthology.org/W04-1013/

    work page 2004

  8. [8]

    In: Annual Meeting of the Association for Computational Linguistics (2004), https://api.semanticscholar.org/CorpusID:964287

    Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Annual Meeting of the Association for Computational Linguistics (2004), https://api.semanticscholar.org/CorpusID:964287

    work page 2004

Show all 31 references
  1. [9]

    CoRRabs/2407.00747(2024), https://arxiv.org/abs/2407.00747

    Nguyen, H., Chen, H., Pobbathi, L., Ding, J.: A comparative study of qual- ity evaluation methods for text summarization. CoRRabs/2407.00747(2024), https://arxiv.org/abs/2407.00747

    2024 arXiv

  2. [10]

    arXiv preprint arXiv:2407.05033 (2024)

    Ramos, J., Wu, B., Lipani, A.: PeaPOD: Personalized prompt distillation for gen- erative recommendation. arXiv preprint arXiv:2407.05033 (2024)

    2024

  3. [11]

    Richardson, C., Zhang, Y., Gillespie, K., Kar, S., Singh, A., Raeesy, Z., Khan, O.Z., Sethy, A.:Integrating summarizationand retrieval forenhancedpersonalization via large language models (2023), https://arxiv.org/abs/2310.20081

    2023 arXiv

  4. [12]

    In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL)

    Salemi, A., Mysore, S., Bendersky, M., Zamani, H.: LaMP: When large language models meet personalization. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). pp. 7370–7392 (2024)

    2024

  5. [13]

    arXiv preprint arXiv:2505.16307 (2025)

    Wang, Z., et al.: PMPO: Probabilistic metric prompt optimization for small and large language models. arXiv preprint arXiv:2505.16307 (2025)

    2025

  6. [14]

    arXiv preprint arXiv:2506.05305 (2025)

    Weerasooriya, T., Liakata, M., et al.: ProRefine: Inference-time prompt refinement with textual feedback. arXiv preprint arXiv:2506.05305 (2025)

    2025

  7. [15]

    In: Proceedings of the Advances in Neural Information Pro- cessing Systems (NeurIPS)

    Yang, C., Wang, X., Lu, Y., Liu, H., Le, Q.V., Zhou, D., Chen, X.: Large language models as optimizers. In: Proceedings of the Advances in Neural Information Pro- cessing Systems (NeurIPS). vol. 36 (2024)

    2024

  8. [16]

    arXiv preprint arXiv:2509.18158 (2025)

    Yi, S., Khang, M., Park, S.: ZERA: Zero-init instruction evolving refinement agent: From zero instructions to structured prompts via principle-based optimization. arXiv preprint arXiv:2509.18158 (2025)

    2025

  9. [17]

    ArXivabs/1904.09675(2019), https://api.semanticscholar.org/CorpusID:127986044

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. ArXivabs/1904.09675(2019), https://api.semanticscholar.org/CorpusID:127986044

    2019 arXiv

  10. [18]

    IEEE Transactions on Knowledge and Data Engineering (2025)

    Zhang, Y., Feng, F., Zhang, J., Bao, K., Wang, Q., He, X.: CoLLM: Integrating collaborative embeddings into large language models for recommendation. IEEE Transactions on Knowledge and Data Engineering (2025)

    2025

  11. [19]

    arXiv preprint arXiv:2303.18223 (2023)

    Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)

    2023 arXiv

  12. [20]

    In: Proceedings of the Inter- national Conference on Learning Representations (ICLR) (2023)

    Zhou, Y., Muresanu, A.I., Han, Z., Paster, K., Pitis, S., Chan, H., Ba, J.: Large language models are human-level prompt engineers. In: Proceedings of the Inter- national Conference on Learning Representations (ICLR) (2023)

    2023

  13. [21]

    Provide summary following the format of reference summary

    Zollo, T.P., Siah, A.W.T., Ye, N., Li, A., Namkoong, H.: PersonalLLM: Tailoring LLMs to individual preferences. In: Proceedings of the International Conference on Learning Representations (ICLR) (2025) A Prompt Templates Appendix A presents representative prompt templates that...

    2025

  14. [22]

    Provide critique inside <Critique></Critique>

  15. [23]

    Provide FULL revised prompt inside <Suggestion></Suggestion>

  16. [24]

    Output only these two sections

  17. [25]

    New prompt must be self-contained

  18. [26]

    Use input-based rules (no reference comparisons)

  19. [27]

    Avoid fallback/default outputs

  20. [28]

    "".strip() A.4 Scorer LLM Template LLM_EVALUATION_PROMPT_TEMPLATE =

    Respect typical target length {policy_block} INPUTS: - Current prompt + score - Generated examples - Optimization history ADDRESS: - Reference-dependent wording - Prefix hallucinations - Content drift - Length mismatch """.strip() A.4 Scorer LLM Template LLM_EVALUATION_PROMPT_...

  21. [29]

    Core Meaning: Semantic correctness and preservation of key information

  22. [30]

    Unsupported Additions: Penalize hallucinated or unsupported content

  23. [31]

    Format & Style Fidelity: Adherence to structural and stylistic constraints Return: Core Meaning: [0-100] Unsupported Additions: [0-100] Format & Style Fidelity: [0-100] Explanation: [...] """.strip()

Pith tools

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