Pith. sign in

REVIEW 3 major objections 5 minor 28 references

No Answer is Better Than Wrong Answer: A Reflection Model for Document Level Machine Reading Comprehension

T0 review · 3 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A two-stage architecture called Reflection Net — a span extractor plus a second model that scores the confidence of the extracted answer — reaches the top of the Natural Questions leaderboard for both long and short answers, with F1…

desk verdict Solid, well-ablated confidence-model paper for NQ; the exact-match training label versus token-level F1 mismatch is a real but non-fatal caveat. read the letter →

arxiv 2009.12056 v2 pith:GUXKZRNV submitted 2020-09-25 cs.CL cs.LG

Xuguang Wang , Linjun Shou , Ming Gong , Nan Duan , Daxin Jiang This is my paper · ORCID
classification cs.CLcs.LG
keywords machinereadingcomprehensionNaturalQuestionsanswerconfidencescoringno-answerdetectionmulti-spanextractiontransformerfine-tuningverificationReflectionNet
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 paper tries to establish that a machine-reading model can be made substantially more reliable when documents contain mixed answer types, especially questions with no answer at all, by adding a second 'Reflection' model that re-scores the first model's prediction. On the Natural Questions benchmark, the combined system reports the top leaderboard F1 of 77.2 for long answers and 64.1 for short answers, beating the previous best by 1.1 and 2.7 points. The reflection step matters because roughly half of Natural Questions are unanswerable from the given Wikipedia page, so a system that confidently emits a wrong answer is worse than one that declines to answer. The paper argues a dedicated confidence model, trained on the first model's own predictions, corrects for the distribution shift between down-sampled training windows and full-document inference, which is why heuristic logit scores end up miscalibrated.

What carries the argument

The Reflection model — a second transformer initialized with the trained MRC model's parameters and fine-tuned to predict whether the first-stage answer is correct (exact string match to the ground truth) — is the mechanism that carries the argument. It takes as input the original context augmented with answer-type and span markers, plus head features (the heuristic score, answer-type probabilities, and top-5 start/end logits and probabilities), and outputs a confidence probability. This confidence score, not the MRC model's heuristic logit score, determines whether the answer is triggered at all; the paper demonstrates this re-ranking is what converts wrong-answer predictions into no-answer or right-answer predictions.

What would settle it

Compute the calibration error (expected calibration error or a reliability diagram) of the Reflection model's confidence scores on the NQ development set, separately for no-answer and answerable questions; if the Reflection probabilities are not better calibrated than the heuristic span scores from Eq. (12), the central claim that reflection 'predicts a more accurate confidence score' is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that a second-stage confidence model — the Reflection model — turns a good span extractor into a substantially better question-answering system on Natural Questions, by learning when the extractor's top answer should be suppressed rather than emitted. The first-stage MRC model, a large pretrained transformer fine-tuned on SQuAD 2.0 and NQ, predicts answer type, the minimal single span for the answer, and sequence labels for multi-span answers, and ranks candidate windows by a heuristic logit score. The Reflection model, initialized from that same MRC model, takes the predicted answer marked in the context, the [CLS] summary representation, and a small head-feature vector (answer-type probabilities, start/end logits and probabilities, and the heuristic score) and outputs a binary correctness probability. Replacing the heuristic score with this probability at answer-triggering time is what the paper credits for reducing wrong-answer emissions: on the dev set it converts hundreds of wrong-answer predictions into no-answer or right-answer predictions, and on the sequestered leaderboard the ensemble reaches F1 77.2 (long) and 64.1 (short), top-ranked at submission time.

Load-bearing premise

The whole approach rests on treating exact string match with the ground-truth answer as the perfect signal for whether a prediction is correct, and on assuming the confidence model trained on the first model's own training-time predictions will generalize to the answers the first model produces at test time.

Editorial extensions

If this is right

  • On the dev set, adding the Reflection model to the RoBERTa-based MRC model improves F1 by 2.9 points for long answers and 3.1 for short answers, gains attributable to the confidence score rather than to better span extraction.
  • The Reflection confidence score delivers large improvements in recall at fixed precision: R@P=90 improves by 12.8 absolute points (long) and 6.8 (short) over the previous top leaderboard entry, meaning fewer hallucinated answers would be surfaced at a production precision bar.
  • Ablating the multi-span and yes/no heads costs 2.2 F1 points on short answers despite these answer types covering only 3.5% and 1% of the data — evidence that systematic answer-type modeling, not just no-answer handling, is part of the gain.
  • The ablation comparison to an answer verifier shows that initializing the second stage from the MRC model and feeding it head features is what makes the Reflection model work: removing both cuts the gain roughly in half, while a plain feedforward net on head features alone beats the heuristic score by 1.1 and 1.7 F1 points.

Reading between the lines

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

  • If the exact-string-match label is the only correctness signal, the same two-stage recipe should transfer to other extractive QA datasets with a substantial no-answer fraction, but the gains would shrink on datasets where correct answers appear in paraphrased surface forms; a testable version would replace exact match with semantic-equivalence labels and compare.
  • The head-feature vector is effectively a hand-crafted summary of the extractor's uncertainty; a natural next step is to ask whether an end-to-end training objective can distill this reflection behavior directly into the first-stage model, collapsing the two stages into one.
  • Because the leaderboard results are dated, a straightforward reproducibility check is to re-run the exact method on the current Natural Questions leaderboard; if the F1 margin over newer systems narrows or reverses, the central contribution is the confidence re-scoring idea rather than an indefinite performance crown.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Reflection Net for the Natural Questions (NQ) benchmark. It first trains an MRC model (BERT/RoBERTa-large) that handles all answer types—no-answer, yes/no, single-span, and multi-span—using sliding windows over documents. It then trains a second 'Reflection' model, initialized from the MRC model's parameters and augmented with hand-crafted head features, to predict whether the MRC model's top-1 predicted answer is correct. The Reflection model's training labels are binary: a predicted answer is labeled correct only if it exactly matches the ground-truth answer string. At inference, the Reflection model's confidence score replaces the MRC model's heuristic logit-based score for answer triggering. Dev-set results show consistent F1 gains when Reflection is added (e.g., RoBERTa-all-type long-answer F1 from 73.0 to 75.9, short-answer from 58.2 to 61.3), and the paper reports top-1 leaderboard F1 scores of 77.2 (long) and 64.1 (short) at the time of writing. Ablations examine answer-type modeling and the contribution of head features and initialization.

Significance. If the results hold, the paper makes a practically useful contribution: a relatively simple two-stage confidence model yields large gains on a benchmark where 51% of questions have no answer, and the R@P=90 improvements (12.8 long, 6.8 short over prior top systems) are substantial. The ablation design in Table 7 is informative, and Table 5 provides a useful error analysis. The paper does not release code and reports single runs for most configurations, which limits reproducibility. More importantly, the Reflection model's supervision uses exact string match, a proxy for the official token-level F1 metric; this mismatch is a genuine correctness-risk concern for the central claim that Reflection provides a 'more accurate confidence score.' The concern is addressable with additional analysis and experiments, so I view it as requiring major revision rather than rejection.

major comments (3)
  1. [Section 2.2, Training Data Generation; Section 3 evaluation metrics] The binary correctness label for Reflection training is defined as 'same to the ground-truth answer' (exact string match), while the official NQ metrics used in Tables 3 and 4 are token-level F1, which awards partial credit for overlapping spans. For multi-span short answers (3.5% of data) and for long answers whose DOM-node boundaries differ from the gold span, exact match is substantially stricter than F1>0, so a predicted answer containing one of several gold spans is labeled incorrect even though it contributes to F1. Because the Reflection model's training target is calibrated to this proxy, the claim that it provides a 'more accurate confidence score' than the MRC heuristic is not directly measured against the optimization metric. I request (i) statistics on the fraction of Reflection training cases where exact-match is 0 but token F1 is >0 (or >0.5), (ii) an analysis of whether the dev gains from Reflection concentrate in exact-match cases, and (iii) at least one variant trained with soft or partial-credit labels (e.g., F1 as a soft target, or a positive label defined as F1>0) to demonstrate robustness. Without this, the supervision signal remains a proxy whose mismatch could in principle suppress partially correct answers and reduce recall at fixed precision.
  2. [Section 3.1 and Tables 3, 7] All non-ensemble results are single training runs with no variance estimates. The differences that drive the conclusions are small relative to typical seed noise: for example, Table 7 shows long-answer F1 differences of 1.2 between the verifier analogue and the full Reflection model, and short-answer differences of 0.7-1.5 among Reflection variants; the leaderboard gains over prior systems are 1.1 and 2.7 F1. I recommend reporting at least three seeds for the main MRC and MRC+Reflection configurations, or bootstrap confidence intervals over the dev set, so that the reader can assess whether the improvements are stable. The absence of code or detailed decoding pseudocode further limits reproducibility.
  3. [Section 2.2, Training Data Generation] The Reflection model's training data are generated by running the trained MRC model over its own training set. Because the MRC model has seen these examples, its predictions there are in-sample and likely more accurate and differently distributed than its predictions on unseen dev or test data. The paper notes a train/predict distribution mismatch caused by negative-window downsampling, but it does not discuss this additional in-sample optimism. A useful control is to generate Reflection training data on a held-out split of the training set (or on the dev set) and compare Reflection performance; if in-sample optimism is large, the current data-generation procedure is overly favorable to the Reflection model's confidence estimates.
minor comments (5)
  1. [Section 2.2, footnote] The footnote contains a typo: 'finial output' should be 'final output'.
  2. [Table 5 and surrounding text] Please specify how the threshold separating 'no-ans' from 'ans' is set for each row. Because MRC and Reflection use different confidence distributions, the optimal threshold may differ, and without this detail the count changes in Table 5 are ambiguous.
  3. [Section 3.1] Please report the number of Reflection training cases and the positive/negative class ratio under exact-match labeling; class balance directly affects calibration and would help readers interpret the binary classifier's behavior.
  4. [Table 1 and general presentation] There are minor formatting issues: 'Montral' should be 'Montréal' in Table 1(c), and several inline equations such as 's ande' lack spacing. These do not affect the technical content.
  5. [Section 1, Introduction] The sentence 'To the best of our knowledge, this is the first work to model all answer types in NQ task' is broad; please clarify whether the claim refers to the full two-stage Reflection system and cite concurrent work on NQ answer types, if any, to avoid overclaiming.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Reflection model is a supervised confidence module evaluated on blind test; self-citations are non-load-bearing.

full rationale

The paper's central claim is an empirical benchmark result on the NQ leaderboard, and its architecture is a two-stage training setup rather than a mathematical derivation that could reduce to its own inputs. The MRC model predicts answer spans and types, and the Reflection model is trained on a binary correctness label derived from whether the MRC model's top-1 predicted answer matches the ground-truth answer (Section 2.2, 'if the predicted answer is same to the ground-truth answer, the label is 1; otherwise 0'). This is a legitimate auxiliary calibration setup, not a self-fulfilling prediction: the correctness labels come from ground truth, and the Reflection model's confidence is evaluated on held-out development data and sequestered test data through the official F1 procedure. The reported F1 improvements are not forced by construction; they depend on whether the learned confidence generalizes. The only self-citations (Li et al. for the B/I/O tagging scheme and Liu et al. as a leaderboard baseline) are not load-bearing: the tagging scheme is standard and the RikiNet comparison is a baseline rather than a supporting premise. The exact-match supervision used for Reflection training versus the token-level F1 evaluation is a calibration mismatch and a possible correctness risk, but it is not circular because nothing in the training objective mathematically entails the reported F1 gains. No circular step can be quoted with an equation-level reduction, so the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

This is an empirical machine learning paper, so the central claim rests on standard supervised learning and shared benchmark assumptions. The main hand-set choices are hyperparameters (e.g., learning rates, top_n features) and the exact-match labeling rule. No new physical or conceptual entities are introduced beyond the Reflection model architecture itself.

free parameters (2)
  • Head feature top_n = 5
    Only the top 5 logits/probs are used as features (Table 2). This truncation is a hand-chosen constant that affects the Reflection model's input.
  • Reflection model learning rate = 5e-6
    Tuned on dev data; listed in Appendix A. This learning rate is several times smaller than the MRC model's, and is central to the Reflection model's fine-tuning behavior.
assumptions (3)
  • domain assumption The Natural Questions ground-truth annotations are correct and complete.
    The paper treats NQ labels as ground truth to compute correctness labels for Reflection training; if labels are noisy, the learned confidence will be biased.
  • domain assumption Exact string match between predicted and ground-truth answers is a valid indicator of correctness.
    Used in Section 2.2 to generate binary labels for the Reflection model; semantically equivalent answers that do not match textually are counted as wrong.
  • domain assumption Pretrained transformer parameters (BERT/RoBERTa) provide a sufficient base representation for both MRC and Reflection models.
    The method builds on pretrained transformers without questioning their coverage; this is standard practice in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of No Answer is Better Than Wrong Answer: A Reflection Model for Document Level Machine Reading Comprehension." pith.science (2026). https://pith.science/paper/GUXKZRNV

@misc{pith2026200912056,
  author       = {Pith},
  title        = {Pith review of: No Answer is Better Than Wrong Answer: A Reflection Model for Document Level Machine Reading Comprehension},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUXKZRNV}},
  note         = {Machine review of arXiv:2009.12056}
}
read the original abstract

The Natural Questions (NQ) benchmark set brings new challenges to Machine Reading Comprehension: the answers are not only at different levels of granularity (long and short), but also of richer types (including no-answer, yes/no, single-span and multi-span). In this paper, we target at this challenge and handle all answer types systematically. In particular, we propose a novel approach called Reflection Net which leverages a two-step training procedure to identify the no-answer and wrong-answer cases. Extensive experiments are conducted to verify the effectiveness of our approach. At the time of paper writing (May.~20,~2020), our approach achieved the top 1 on both long and short answer leaderboard, with F1 scores of 77.2 and 64.1, respectively.

Figures

Figures reproduced from arXiv: 2009.12056 by the authors.

Figure 1
Figure 1. Overview of our proposed Reflection Net, consisting of MRC model and its corresponding Reflection model. MRC model try its best to predict answer, Reflection model output corresponding answer confidence score. The left arrow denotes when training, Reflection model is initialized with the parameters of trained MRC model. ous approaches have proved effective to improve the performance on the NQ task, few works focus o… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Chris Alberti, Kenton Lee, and Michael Collins. 2019. A bert baseline for the natural questions. arXiv preprint arXiv:1901.08634

    work page Pith review arXiv 2019

  4. [4]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. http://arxiv.org/abs/1409.0473 Neural machine translation by jointly learning to align and translate . In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    work page Pith review arXiv 2015

  5. [5]

    Longformer: The Long-Document Transformer

    Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. ArXiv, abs/2004.05150

    work page Pith review arXiv 2020

  6. [6]

    Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. https://doi.org/10.18653/v1/P17-1171 Reading W ikipedia to answer open-domain questions . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870--1879, Vancouver, Canada. Association for Computational Linguistics

    work page doi:10.18653/v1/p17-1171 2017

  7. [7]

    Christopher Clark and Matt Gardner. 2018. https://doi.org/10.18653/v1/P18-1078 Simple and effective multi-paragraph reading comprehension . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 845--855, Melbourne, Australia. Association for Computational Linguistics

    work page doi:10.18653/v1/p18-1078 2018

  8. [8]

    Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. 2017. https://doi.org/10.18653/v1/P17-1055 Attention-over-attention neural networks for reading comprehension . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 593--602, Vancouver, Canada. Association for Compu...

    work page doi:10.18653/v1/p17-1055 2017

Show all 28 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  2. [10]

    Qipeng Guo, Xipeng Qiu, Pengfei Liu, Yunfan Shao, Xiangyang Xue, and Zheng Zhang. 2019. https://doi.org/10.18653/v1/N19-1133 Star-transformer . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language...

    2019 doi

  3. [11]

    Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415

    2016 arXiv

  4. [12]

    Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in neural information processing systems, pages 1693--1701

    2015

  5. [13]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531

    2015 arXiv

  6. [14]

    Minghao Hu, Furu Wei, Yuxing Peng, Zhen Huang, Nan Yang, and Dongsheng Li. 2019. Read+ verify: Machine reading comprehension with unanswerable questions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6529--6537

    2019

  7. [15]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. http://arxiv.org/abs/1412.6980 Adam: A method for stochastic optimization . In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015 arXiv

  8. [16]

    Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. https://openreview.net/forum?id=rkgNKkHtvB Reformer: The efficient transformer . In International Conference on Learning Representations

    2020

  9. [17]

    Souvik Kundu and Hwee Tou Ng. 2018. https://doi.org/10.18653/v1/D18-1456 A nil-aware answer extraction framework for question answering . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4243--4252, Brussels, Belgium. Association...

    2018 doi

  10. [18]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019 doi

  11. [19]

    Peng Li, Wei Li, Zhengyan He, Xuguang Wang, Ying Cao, Jie Zhou, and Wei Xu. 2016. Dataset and neural recurrent sequence labeling model for open-domain factoid question answering. arXiv preprint arXiv:1607.06275

    2016 arXiv

  12. [20]

    Dayiheng Liu, Yeyun Gong, Jie Fu, Yu Yan, Jiusheng Chen, Daxin Jiang, Jiancheng Lv, and Nan Duan. 2020. https://doi.org/10.18653/v1/2020.acl-main.604 R iki N et: Reading W ikipedia pages for natural question answering . In Proceedings of the 58th Annual Meeting of the Associat...

    2020 doi

  13. [21]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  14. [22]

    Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. https://doi.org/10.18653/v1/P18-2124 Know what you don ' t know: Unanswerable questions for SQ u AD . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages ...

    2018 doi

  15. [23]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. https://doi.org/10.18653/v1/D16-1264 SQ u AD : 100,000+ questions for machine comprehension of text . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-...

    2016 doi

  16. [24]

    Chuanqi Tan, Furu Wei, Qingyu Zhou, Nan Yang, Weifeng Lv, and Ming Zhou. 2018. I know there is no answer: Modeling answer validation for machine reading comprehension. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 85--97. Springer

    2018

  17. [25]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998--6008

    2017

  18. [26]

    Wei Wang, Ming Yan, and Chen Wu. 2018. https://doi.org/10.18653/v1/P18-1158 Multi-granularity hierarchical attention fusion networks for reading comprehension and question answering . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (V...

    2018 doi

  19. [27]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, et al. 2019. Transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771

    2019 arXiv

  20. [28]

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237

    2019 arXiv

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.