Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Temporalizing Confidence: Evaluation of Chain-of-Thought Reasoning with Signal Temporal Logic

T0 review · 4 major / 5 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 argues that stepwise chain-of-thought confidence, treated as a temporal signal and scored with signal temporal logic robustness, yields better-calibrated LLM uncertainty estimates than conventional aggregation or post-hoc…

desk verdict A reasonable first pass at temporal-logic confidence for chain-of-thought, but the missing extraction procedure and missing non-temporal control mean the headline claim is not yet supported. read the letter →

arxiv 2506.08243 v1 pith:2UDZA7K5 submitted 2025-06-09 cs.LG cs.AI

Zhenjiang Mao , Artem Bisliouk , Rohith Reddy Nama , Ivan Ruchkin This is my paper
classification cs.LGcs.AI
keywords chain-of-thoughtreasoningconfidencecalibrationsignaltemporallogicuncertaintyestimationexpectederrorBrierscoremathematical
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 the step-by-step confidence an LLM expresses while reasoning contains usable signal that conventional single-number confidence estimates throw away. It models the sequence of per-step confidence values as a time-varying signal, smooths it with four reshaping rules, and scores the result against formal temporal constraints—Eventually Confident, Always Stable, Locally Smooth—using signal temporal logic robustness. On 1,296 Gaokao math multiple-choice items, these STL-based scores yield lower expected calibration error and Brier scores than averaging, temperature scaling, or histogram binning, with the best ECE dropping to 0.056. If right, the paper offers a post-hoc, model-agnostic way to turn chain-of-thought traces into more trustworthy confidence estimates without retraining the model.

What carries the argument

The central mechanism is the Signal Temporal Logic (STL) robustness score, a formal-language tool that measures how far a real-valued time signal satisfies a temporal specification: a positive score gives the satisfaction margin, a negative score gives the violation magnitude. The paper encodes three specifications—Eventually Confident, Always Stable, and Locally Smooth—and uses their ReLU-transformed robustness values as interpretable confidence estimates. The reshaping strategies (CMS, EDS, MPS, GS) are the supporting machinery: they smooth the raw per-step confidence signal before STL evaluation so that the logic scores reflect temporally coherent trajectories rather than noisy spikes.

What would settle it

Shuffle the per-step confidence values within every reasoning trace before applying reshaping and STL scoring, then recompute ECE and Brier Score. If the calibration gains survive shuffling, temporal order is not the carrier; if they vanish, the trajectory shape is doing the work.

Watch

Extended reading notes

Core claim

The central claim is that the temporal shape of stepwise confidence carries information about the correctness of an LLM's final answer, and that this information can be extracted without retraining. The paper defines four reshaping strategies—Causal Minimum Smoothing, Exponential Decay Smoothing, Monotonic Penalty Smoothing, and Guarded Smoothing—that suppress abrupt and causally suspicious confidence spikes. The reshaped trajectory is then evaluated against three signal temporal logic formulas, each capturing a normative property of good reasoning: confidence should eventually rise above a threshold, should not drop abruptly, and should not change too much between steps. The robustness score of each formula, passed through ReLU to land in $[0,1]$, becomes the confidence estimate. Across logits-based, self-evaluation-based, and internal analysis-based confidence sources, these STL-derived scores consistently improve Expected Calibration Error and Brier Score relative to standard post-hoc calibration methods, with the best internal-based ECE at 0.056 using Always Stable combined with Exponential Decay Smoothing.

Load-bearing premise

The load-bearing premise is that every reasoning step can be assigned a meaningful scalar confidence value and that the ordering of those values over time carries information about whether the final answer is correct; the paper states each step has such a score but does not specify how it is extracted.

Editorial extensions

If this is right

  • STL-derived confidence scores can be computed post-hoc from any model's chain-of-thought trace, without retraining or modifying the model.
  • Combining a smoothing strategy with an STL constraint—for example Always Stable with Exponential Decay Smoothing—achieves the lowest calibration error on this benchmark, down to an ECE of 0.056.
  • The scores double as temporal diagnostics, indicating whether a reasoning trajectory eventually becomes confident, remains stable, and changes smoothly.
  • The approach transfers to any multiple-choice or classification setting in which the model emits stepwise reasoning and a scalar confidence per step.
  • Lower Brier Scores accompany the ECE improvements, so the gains reflect sharper probability estimates, not just better ordering of predictions.

Reading between the lines

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

  • If the temporal-shape signal is real, the same pipeline could localize the reasoning step where confidence first becomes misleading by evaluating STL robustness over sub-trajectories.
  • A testable consequence is that the calibration gain shrinks on models whose raw confidence is already well calibrated; the framework should matter most where overconfidence is severe.
  • Part of the ECE gain may come from the reshaping rules' downward bias on confidence values rather than from the temporal logic itself; an ablation replacing STL scoring with a time-averaged statistic after the same reshaping would separate these mechanisms.
  • The paper's stated limitation about branching reasoning suggests a natural extension: tree-structured confidence could be evaluated with branching-time logic, an avenue the authors explicitly flag as future work.
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 framework for estimating confidence in chain-of-thought (CoT) reasoning by treating the per-step confidence values as a temporal signal, applying reshaping strategies (CMS, EDS, MPS, GS), and computing Signal Temporal Logic (STL) robustness scores (STL1--STL3) as structured confidence estimates. The method is evaluated on GAOKAO-Bench multiple-choice math questions using Qwen-7B, reporting expected calibration error (ECE) and Brier scores across three confidence sources (logits, self-evaluation, internal). The authors claim consistent improvements over conventional aggregation and post-hoc calibration baselines.

Significance. If the central claim were fully supported, the paper would introduce a genuinely novel perspective: using formal temporal specifications to evaluate the structure of reasoning confidence, which could provide interpretable diagnostics beyond scalar calibration metrics. The STL formulations are simple and could inspire further work on logic-based uncertainty quantification. The paper also provides a useful ablation of reshaping strategies and evaluates on a realistic educational benchmark. However, the current evidence is insufficient to establish that STL-based temporal evaluation, rather than the reshaping transformations or threshold tuning, is responsible for the reported gains.

major comments (4)
  1. [§3.2] The load-bearing assumption that each CoT step is 'associated with a confidence score c_t' is never operationalized. No extraction procedure is described for any of the three confidence sources: there is no prompt template, no per-token or per-step logit aggregation formula, no verbalized-probability protocol, and no algorithm for the 'internal-based' score. Since CMS, EDS, MPS, GS, and all STL robustness values are deterministic functions of the c_t sequence, the entire evaluation is contingent on an unstated choice. The authors must specify exactly how c_t is produced for each source, or the experimental results cannot be reproduced or meaningfully interpreted.
  2. [Table 2, §4] The claim that 'STL-based temporal evaluation further improves calibration' is not isolated from the reshaping step. The table shows that reshaping alone (e.g., logits CMS ECE = 0.107) already outperforms several STL-based configurations (e.g., logits STL1-only = 0.174, STL1+CMS = 0.250, STL1+EDS = 0.236). To attribute gains to temporal structure, the authors need a control that applies the same reshaping and then uses a non-temporal statistic over the same reshaped signal, such as the mean, median, or a monotone transform of the minimum. Without this control, the improvement could equally be attributed to the reshaping transformations alone.
  3. [§4, Tables 2 and 3] All results are reported over only three runs, with no significance tests or confidence intervals on the differences between methods. Several standard deviations are very large relative to the reported improvements (e.g., STL1+CMS logits ECE = 0.250 ± 0.198; STL1+EDS internal = 0.500 ± 0.497). The abstract and Section 4 state that the approach 'consistently improves' calibration, but a 0.500 ± 0.497 result is statistically indistinguishable from chance and directly undermines that wording. The authors should report paired significance tests across the three runs or otherwise quantify the reliability of the rank ordering.
  4. [§3.3, §4] The method is described as 'partially post-hoc' because the thresholds τ, ε, δ are grid-searched on a validation set. This is acknowledged in the text, but the comparison is still unfair: baselines such as temperature scaling and histogram binning are also tuned, yet no non-temporal calibration method is given the same budget of reshaping plus tuned threshold. Additionally, Figure 3, which is used to illustrate the benefit of STL, is said to reflect 'a single problem instance and one STL constraint,' so it cannot support the aggregate claim of improvement. The authors should provide a fair comparison where the non-temporal control uses the same tuned thresholds and reshaping, and should either remove Figure 3's implied generality or explicitly label it as a case study.
minor comments (5)
  1. [§3.1] There is an incomplete sentence after Equation (2): 'rect, and 0 otherwise.' appears to be a leftover fragment. It should be removed or completed.
  2. [§4, paragraph 1] The first sentence of Section 4 contains duplicated and tangled phrasing: 'In Section 4, we present an ablation study and a comparison against established post-hoc calibration techniques that investigates the impact of STL parameterization and compares our method against established post-hoc calibration techniques such as Temperature Scaling...' The sentence should be reworded for clarity.
  3. [§4, text after Table 2] The text states that 'STL1 combined with GS achieves an ECE of 0.077 on self-evaluation-based confidence and 0.096 on internal-based, outperforming all other approaches.' However, Table 2 shows STL3 with EDS achieves self-evaluation ECE = 0.076, which is slightly better, so the claim of outperforming all other approaches is inaccurate for that cell.
  4. [§3.2, bullet definitions] The reshaping definitions use τ in MPS and GS and also in STL1, but the relationships among these thresholds are not discussed. For instance, is the τ in GS the same as the τ in STL1? Clarifying the reuse of symbols would help reproducibility.
  5. [Abstract and §1] The abstract claims the approach 'consistently improves calibration metrics,' but Tables 2 and 3 show many configurations where STL-based scores are worse than simple CoT averaging (e.g., logits STL1-only ECE = 0.174 vs. CoT average 0.141; logits STL1+CMS 0.250 vs. CMS 0.107). The wording should be softened to reflect the specific configurations that improve.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the STL robustness pipeline is an explicit post-hoc transform with validation-tuned thresholds and held-out test evaluation.

full rationale

The derivation chain is not circular. Stepwise confidence c_t is taken as input (Section 3.2); each reshaping function (CMS, EDS, MPS, GS) is an explicit arithmetic transform; each STL robustness score is the standard quantitative semantics definition from the external STL literature (Fainekos and Pappas, Donzé and Maler); and the final confidence estimate is ReLU of the robustness value. No equation defines an input in terms of the target calibration output, and no STL formula is derived from the ECE or Brier metric. The thresholds tau, epsilon, and delta are grid-searched on a held-out validation set, and all reported ECE and Brier values are on the test set, so this is ordinary hyperparameter fitting rather than a fitted parameter being renamed as a prediction. The comparison baselines (temperature scaling, histogram binning) are likewise fit on validation data. The paper explicitly acknowledges that the STL evaluation stage is "partially post-hoc" and requires data-driven tuning. No load-bearing self-citation, uniqueness theorem, or imported ansatz is used: the STL and calibration references are external prior work. The under-specified extraction of per-step confidence is a methodological gap, not a circular step, because the subsequent reshaping and STL scoring are well-defined functions of whatever c_t sequence is supplied. The central claim therefore has independent empirical content and is testable against the held-out Gaokao data.

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

No novel physical entities are introduced. The main assumptions are domain assumptions about stepwise confidence signals and dataset augmentation; the only free numbers are thresholds and smoothing weights tuned on validation data.

free parameters (4)
  • tau (STL and reshaping threshold) = grid-searched on validation set, value not reported
    Used in STL1 (eventually c(t) > tau), MPS threshold, and GS cap; chosen by validation grid search.
  • epsilon (STL2 and GS tolerance) = grid-searched on validation set, value not reported
    Used in STL2 (always stable) and Guarded Smoothing; fitted.
  • delta (CMS margin and STL3 bound) = grid-searched on validation set, value not reported
    Used in Causal Minimum Smoothing margin and STL3 local smoothness bound; fitted.
  • alpha (EDS smoothing weight) = not reported
    Exponential Decay Smoothing blending weight; the paper does not state how alpha is chosen.
assumptions (4)
  • domain assumption Each chain-of-thought step has an associated scalar confidence c_t that is meaningful and extractable
    Section 3.2 states 'each associated with a confidence score c_t' but gives no extraction procedure; the whole pipeline depends on this signal.
  • domain assumption Smooth, monotonic, causally consistent confidence evolution is desirable and correlated with correctness
    Section 3.2 motivates reshaping by claiming abrupt confidence increases after low-confidence steps are misleading; no empirical or theoretical support is given for this mapping.
  • domain assumption Paraphrased and backtranslated versions of Gaokao questions preserve logical structure and the correct answer
    Section 4 says answers are inherited from original problems because paraphrasing preserves equivalence, with no verification of equivalence.
  • standard math STL robustness semantics provide a valid quantitative satisfaction measure for finite discrete signals
    Background from Fainekos and Pappas 2006 and Donze and Maler 2010, accepted formalism.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Temporalizing Confidence: Evaluation of Chain-of-Thought Reasoning with Signal Temporal Logic." pith.science (2026). https://pith.science/paper/2UDZA7K5

@misc{pith2026250608243,
  author       = {Pith},
  title        = {Pith review of: Temporalizing Confidence: Evaluation of Chain-of-Thought Reasoning with Signal Temporal Logic},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2UDZA7K5}},
  note         = {Machine review of arXiv:2506.08243}
}
read the original abstract

Large Language Models (LLMs) have shown impressive performance in mathematical reasoning tasks when guided by Chain-of-Thought (CoT) prompting. However, they tend to produce highly confident yet incorrect outputs, which poses significant risks in domains like education, where users may lack the expertise to assess reasoning steps. To address this, we propose a structured framework that models stepwise confidence as a temporal signal and evaluates it using Signal Temporal Logic (STL). In particular, we define formal STL-based constraints to capture desirable temporal properties and compute robustness scores that serve as structured, interpretable confidence estimates. Our approach also introduces a set of uncertainty reshaping strategies to enforce smoothness, monotonicity, and causal consistency across the reasoning trajectory. Experiments show that our approach consistently improves calibration metrics and provides more reliable uncertainty estimates than conventional confidence aggregation and post-hoc calibration.

Figures

Figures reproduced from arXiv: 2506.08243 by the authors.

Figure 1
Figure 1. (a) Conventional methods output global confidence via logit, self-evaluation, or internal analysis. (b) Our [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization examples of stepwise confidence signals before and after applying different uncertainty [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. ECE comparison of four confidence estimation methods. (1) Final-step confidence; (2) average confidence [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Systematic Optimization of Open Source Large Language Models for Mathematical Reasoning

    cs.LG 2025-09 reject novelty 3.0 of 10

    A hyperparameter search for LLM math reasoning that reports simulated, not measured, performance gains.

Reference graph

Works this paper leans on

32 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    Trista M Ben \' tez, Yueyuan Xu, J Donald Boudreau, Alfred Wei Chieh Kow, Fernando Bello, Le Van Phuoc, Xiaofei Wang, Xiaodong Sun, Gilberto Ka-Kit Leung, Yanyan Lan, et al. 2024. Harnessing the potential of large language models in medical education: promise and pitfalls. Journal of the American Medical Informatics Association, 31(3):776--783

    work page 2024

  2. [2]

    Shrey Desai and Greg Durrett. 2020. Calibration of pre-trained transformers. arXiv preprint arXiv:2003.07892

    arXiv 2020

  3. [3]

    Shizhe Diao, Pengcheng Wang, Yong Lin, Rui Pan, Xiang Liu, and Tong Zhang. 2023. Active prompting with chain-of-thought for large language models. arXiv preprint arXiv:2302.12246

    arXiv 2023

  4. [4]

    Alexandre Donz \'e and Oded Maler. 2010. Robust satisfaction of temporal logic over real-valued signals. In International Conference on Formal Modeling and Analysis of Timed Systems, pages 92--106. Springer

    work page 2010

  5. [5]

    Georgios E Fainekos and George J Pappas. 2006. Robustness of temporal logic specifications. In International Workshop on Formal Approaches to Software Testing, pages 178--192. Springer

    work page 2006

  6. [6]

    Wensheng Gan, Zhenlian Qi, Jiayang Wu, and Jerry Chun-Wei Lin. 2023. Large language models in education: Vision and opportunities. In 2023 IEEE international conference on big data (BigData), pages 4776--4785. IEEE

    work page 2023

  7. [7]

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321--1330. PMLR

    work page 2017

  8. [8]

    Mohamed Diab Idris, Xiaohua Feng, and Vladimir Dyo. 2024. Revolutionising higher education: Unleashing the potential of large language models for strategic transformation. IEEE Access

    work page 2024

Show all 32 references
  1. [9]

    Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2021. How can we know when language models know? on the calibration of language models for question answering. Transactions of the Association for Computational Linguistics, 9:962--977

    2021

  2. [10]

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221

    2022 arXiv

  3. [11]

    u chemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan G \

    Enkelejda Kasneci, Kathrin Se ler, Stefan K \"u chemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan G \"u nnemann, Eyke H \"u llermeier, et al. 2023. Chatgpt for good? on opportunities and challenges of large language models for education....

    2023

  4. [12]

    Xiao Li, Cristian-Ioan Vasile, and Calin Belta. 2017. Reinforcement learning with temporal logic rewards. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3834--3839. IEEE

    2017

  5. [13]

    Qing Lyu, Kumar Shridhar, Chaitanya Malaviya, Li Zhang, Yanai Elazar, Niket Tandon, Marianna Apidianaki, Mrinmaya Sachan, and Chris Callison-Burch. 2025. Calibrating large language models with sample consistency. In Proceedings of the AAAI Conference on Artificial Intelligence...

    2025

  6. [14]

    Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. 2024. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229

    2024 arXiv

  7. [15]

    Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. 2015. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, volume 29

    2015

  8. [16]

    Nikoleta Polyxeni Paulina Kastania. 2024. Building trust in ai education: Addressing transparency and ensuring. Trust and Inclusion in AI-mediated Education: Where Human Learning Meets Learning Machines, page 73

    2024

  9. [17]

    Nicholas Rescher and Alasdair Urquhart. 2012. Temporal logic, volume 3. Springer Science & Business Media

    2012

  10. [18]

    Vedant Shah, Dingli Yu, Kaifeng Lyu, Simon Park, Jiatong Yu, Yinghui He, Nan Rosemary Ke, Michael Mozer, Yoshua Bengio, Sanjeev Arora, et al. 2024. Ai-assisted generation of difficult math questions. arXiv preprint arXiv:2407.21009

    2024 arXiv

  11. [19]

    Sree Harsha Tanneru, Chirag Agarwal, and Himabindu Lakkaraju. 2024. Quantifying uncertainty in natural language explanations of large language models. In International Conference on Artificial Intelligence and Statistics, pages 1072--1080. PMLR

    2024

  12. [20]

    Birgit C van Huijgevoort, Ruohan Wang, Sadegh Soudjani, and Sofie Haesaert. 2024. Specification-guided temporal logic control for stochastic systems: a multi-layered approach. arXiv preprint arXiv:2407.03896

    2024 arXiv

  13. [21]

    Shen Wang, Tianlong Xu, Hang Li, Chaoli Zhang, Joleen Liang, Jiliang Tang, Philip S Yu, and Qingsong Wen. 2024. Large language models for education: A survey and outlook. arXiv preprint arXiv:2403.18105

    2024 arXiv

  14. [22]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171

    2022 arXiv

  15. [23]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  16. [24]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809--11822

    2023

  17. [25]

    Bianca Zadrozny and Charles Elkan. 2001. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. In Icml, volume 1

    2001

  18. [26]

    Xiaotian Zhang, Chunyang Li, Yi Zong, Zhengyu Ying, Liang He, and Xipeng Qiu. 2023. Evaluating the performance of large language models on gaokao benchmark. arXiv preprint arXiv:2305.12474

    2023 arXiv

  19. [27]

    Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2022. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493

    2022 arXiv

  20. [28]

    Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International conference on machine learning, pages 12697--12706. PMLR

    2021

  21. [29]

    Yizhou Zhou, Mengqiao Zhang, Yuan-Hao Jiang, Xinyu Gao, Naijie Liu, and Bo Jiang. 2025. A study on educational data analysis and personalized feedback report generation based on tags and chatgpt. arXiv preprint arXiv:2501.06819

    2025 arXiv

  22. [30]

    Yuqi Zhu, Ge Li, Xue Jiang, Jia Li, Hong Mei, Zhi Jin, and Yihong Dong. 2025. Uncertainty-guided chain-of-thought for code generation with llms. arXiv preprint arXiv:2503.15341

    2025 arXiv

  23. [31]

    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...

  24. [32]

    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 gl...

Pith tools

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