Pith. sign in

REVIEW 5 major objections 5 minor 29 references

Respecting Temporal-Causal Consistency: Entity-Event Knowledge Graphs for Retrieval-Augmented Generation

T0 review · 5 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 claims that preserving entity mentions as distinct nodes and linking each mention to its events in a bipartite entity-event graph lets retrieval-augmented generation answer temporal and causal narrative questions better than…

desk verdict ChronoQA is a genuinely useful benchmark and E2RAG is a plausible idea, but the paper's central structural claim is currently supported only by an interaction with HyDE, not by a head-to-head test of the graph. read the letter →

arxiv 2506.05939 v1 pith:M3SCESE6 submitted 2025-06-06 cs.IR

Ze Yu Zhang , Zitao Li , Yaliang Li , Bolin Ding , Bryan Kian Hsiang Low This is my paper
classification cs.IR
keywords retrieval-augmentedgenerationentity-eventknowledgegraphtemporalreasoningcausalconsistencycharacternarrativeQAbenchmarkRAGhypotheticalresponse
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 retrieval-augmented generation (RAG) fails on narrative questions that depend on when an event happened relative to another, and that the failure is fixable by changing how the retriever's graph is built. The proposed system keeps every entity mention as a separate node and links each mention to the event snippets in which it appears, so an early-story character and the same character later in the story remain distinct. On a new benchmark of 497 questions over nine public-domain stories, the best variant of this entity-event graph scores higher on average than unstructured RAG, single-node knowledge-graph RAG, and hypothetical-response baselines, with the largest margins on causal and character-consistency questions. If the result holds, the paper offers a drop-in way to make RAG respect chronological context without fine-tuning the generator.

What carries the argument

The load-bearing object is the entity-event knowledge graph: an entity subgraph $G_{\mathrm{ent}}$, an event subgraph $G_{\mathrm{evt}}$, and a bipartite edge set $B = \{(v_i, e_j) \mid v_i \in V_{\mathrm{ent}}, e_j \in V_{\mathrm{evt}}, \mathrm{Name}(v_i) \subset \mathrm{Desc}(e_j)\}$ that ties each entity mention to the events in its own chunk. What does the work at query time is the one-hop expansion over $B$ followed by event-grounded re-ranking: an entity candidate is promoted only when its paired event nodes also match the query, which is how chronological context survives retrieval. A second component is the hypothetical-response coupling, a single extra LLM call that drafts a prospective answer and feeds it, together with the original query, into the entity and event cue extractor to enrich the embedding.

What would settle it

Take the 497 ChronoQA answers and re-score every system's response against the ground-truth passage offsets with one deterministic rubric, such as checking whether all entity and event facts in the answer match the cited excerpt, then compute bootstrap confidence intervals on the mean; if the best E2RAG variant no longer separates from the best baseline, the claimed temporal-causal advantage is not established.

Watch

Extended reading notes

Core claim

The paper's central claim is that the collapse of entity mentions into one node is not a harmless compression: it erases exactly the temporal and causal information needed for narrative reasoning. To test this, the authors built ChronoQA, a passage-grounded benchmark of 497 question-answer pairs over nine public-domain narratives, and E2RAG, a dual-graph retriever that keeps an entity-mention subgraph and an event subgraph joined by a bipartite edge set. Retrieval starts from LLM-extracted entity and event cues, pulls seed nodes from two vector stores, expands one hop across the bipartite edges so every retrieved entity arrives with its time-specific events, and re-ranks entity nodes by the similarity of their associated events. The best variant, which feeds a hypothetical answer and the original query together into the extractor, reaches an average judge score of 7.125 on ChronoQA, the highest in the study, and an E2RAG variant is the top scorer in every reasoning category.

Load-bearing premise

The load-bearing premise is that the mean of three LLM judges' 1-10 Likert scores is a trustworthy measure of answer quality; the judges differ in average calibration by nearly four points on the same top system, so noisy or lenient judging could change the ranking.

Editorial extensions

If this is right

  • Narrative-focused RAG systems can adopt a mention-level entity-event index as a drop-in replacement for single-node entity graphs without fine-tuning the generator.
  • Feeding a hypothetical answer together with the original query into the cue extractor, a single extra LLM call, is enough to yield the largest observed gains.
  • ChronoQA provides public, passage-offset-grounded questions that later retrieval systems can be measured against directly.
  • Because every reasoning category is topped by an E2RAG variant, the benefit extends past causal and character queries to symbolic, thematic, and plot-structure questions.
  • The paper states the framework targets documents with inherent temporal or causal structure and may not help on corpora lacking such structure.

Reading between the lines

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

  • A natural ablation would remove the bipartite edge set $B$ and run the same retrieval; if the score drops, the event grounding itself, not merely the extra node count, is the active ingredient.
  • The same mention-preserving design could transfer to timelines outside fiction, such as biographies, medical histories, or news stories, wherever an entity's state changes over time; the paper does not test these settings.
  • Because two of the three judges calibrate very differently, re-scoring the released answers with a single judge or a deterministic fact-check would tell whether the roughly quarter-point margin over the best baseline is stable; the paper reports no significance test.
  • A cheaper preprocessing variant could build $B$ by co-occurrence of entity and event spans in the same chunk instead of scanning for entity names inside event descriptions, cutting the token cost that the paper notes doubles when event extraction is added.
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

5 major / 5 minor

Summary. The paper introduces ChronoQA, a 497-question retrieval QA benchmark over nine public-domain narrative documents, with byte-offset passage evidence for each answer. It also proposes E2RAG, a dual-graph retrieval framework that keeps entity mentions as separate nodes and links them to event nodes through a bipartite edge set B, followed by seed retrieval and one-hop expansion; four HyDE-based query variants are defined. On ChronoQA, the best variant (E2RAG combined extraction) obtains an average judge score of 7.125, ahead of LightRAG hybrid (6.880) and GraphRAG drift (6.820), while E2RAG without hypothetical responses scores 6.708 and ranks eighth. The paper claims notable gains on causal and character-consistency queries and positions E2RAG as a drop-in upgrade for narrative RAG.

Significance. ChronoQA is a useful community resource: it is openly released, built from public-domain texts, and provides passage-level supervision, which existing narrative long-context benchmarks largely lack. The E2RAG design is elegant and the paper includes a concrete case study, preprocessing and query-time cost comparisons, and pseudocode for the construction and retrieval algorithms. The central empirical claim, however, is not yet supported. The evaluation never measures retrieval quality directly, the vanilla dual-graph system is not competitive with the strongest baselines, and the reported leaderboard margins are within plausible judge noise. With retrieval metrics and a cleaner ablation, the paper could substantiate its claim; in its current form the evidence is suggestive rather than conclusive.

major comments (5)
  1. [Section 3 and Section 5 (Tables 3-21)] Despite ChronoQA providing byte-offset ground-truth passages, no retrieval-stage metric (Recall@k, MRR, hit rate) is reported against those passages. Since the paper's central claim is that the dual graph preserves temporal-causal context for retrieval, the absence of a direct retrieval evaluation means the mechanism is never tested independently of the downstream generator and judge. This is load-bearing: the passage evidence exists precisely to make such verification straightforward.
  2. [Section 5, Tables 3 and 15] E2RAG (vanilla), the variant that contains the dual graph but no hypothetical-response augmentation, scores 6.708 overall and 6.910 on Character Consistency, below LightRAG hybrid (6.880 and 7.121) and GraphRAG drift (6.820 and 6.975). The four E2RAG variants that occupy the top four overall slots all incorporate HyDE, while the same HyDE variants applied to LightRAG hybrid drop it from 6.880 to roughly 5.7. Consequently, the reported gain is an interaction between the E2RAG index and hypothetical-response query expansion, not evidence that the dual-graph structure alone improves retrieval. Please report retrieval metrics and an ablation that isolates the graph contribution (e.g., E2RAG vanilla vs. a HyDE-only retriever without the graph) with significance tests.
  3. [Section 5, Eq. (2), Appendix G Tables 11-13] The ranking is based on the mean of three LLM judge scores with no confidence intervals or significance tests. Per-judge rankings disagree: GPT-4o ranks LightRAG hybrid first (6.917) while the other two judges rank E2RAG (comb. extraction) first; Claude-3.7-Sonnet gives the top system a mean of 5.356 and GPT-4.1-mini gives it 9.163. The 0.245 margin between the first and fifth systems is small relative to this inter-judge calibration spread, so the headline ordering may be an artifact of judge composition. Paired significance tests (e.g., bootstrap or Wilcoxon over the 497 items) and per-category error bars should be reported.
  4. [Section 4, Eq. (1), Algorithm 2, Step 3] The formal definition of the bipartite edge set B omits the same-chunk condition that Algorithm 2 enforces ("with CHUNK ID(e) = c"). As written, Eq. (1) connects any entity mention to any event whose description contains the entity name, even across chunks, which would allow edges between entity mentions and events from different chunks, contradicting the algorithm and potentially changing retrieval behavior. The definition and the algorithm must be reconciled.
  5. [Section 5, Tables 14-21] Several category tables are based on very small numbers of questions (e.g., Emotional & Psychological has 2 questions per Table 2), yet the text states that a top-scoring E2RAG mode appears in every category. Such small-sample rankings carry little information and should be omitted or clearly flagged, and the category-level claims should be qualified by the absence of significance tests.
minor comments (5)
  1. [Throughout] There are several typos: "requried" (Section 3), "variance" where "variant" is meant (Section 4), "the them" (Section 4), and "invoved" (Appendix D.1).
  2. [Appendix C] The appendix says hyperparameters were varied with no noticeable benefit, but no details or numbers are given; please report the ranges tried and the resulting scores.
  3. [Table 6 and Appendix F] The case study is based on a single query; multiple examples or a quantitative retrieval analysis would make the qualitative claim more convincing.
  4. [Section 5, Baselines] The paper does not specify how the hypothetical-response variants were applied to LightRAG hybrid (e.g., the same HyDE prompt and the same embedding combination). Please provide these details to rule out implementation differences as an explanation for the poor LightRAG hybrid HyDE results.
  5. [Appendix D.2] The scoring rubric anchors only at 10, 7, 5, 3, and 1, yet scores like 6.6667 and 6.9096 are reported; please clarify how intermediate scores are defined or whether judges produce fractional scores.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: E2RAG's reported gains are empirical results rather than consequences of its construction, and no equation or fitted parameter is defined in terms of the benchmark answers.

full rationale

The paper's central claim is that the E2RAG dual-graph retrieval mechanism improves RAG answer quality on temporal/causal narrative queries. This claim is supported by end-to-end experiments on the new ChronoQA benchmark, not by a derivation that reduces to the method's own definitions. The bipartite edge set in Eq. (1) is defined by string containment (Name(vi) ⊂ Desc(ej)), and the retrieval algorithm in Section 4 is parameter-free apart from chunk size and top-k, with no parameter fitted to ChronoQA answers or ground-truth passages. Consequently, there is no statistical or definitional forcing of the reported scores: E2RAG's ranking in Table 3 is an empirical outcome, not a tautology. The concerns raised by the skeptic are validity and attribution issues rather than circularity: the benchmark is generated by LLM oracles, the judges are also LLMs with substantially different calibration (Appendix G), and the top E2RAG variants all incorporate HyDE, so the source of the gains is not isolated. These are legitimate experimental-design concerns but do not make any prediction equivalent to its input by construction. The mismatch between Eq. (1), which omits the same-chunk restriction, and Algorithm 2, which includes it, is an internal inconsistency in the formal definition, not a circular step. There are no load-bearing self-citations: the references to LightRAG, GraphRAG, and HyDE are external prior work, and no uniqueness theorem or prior result by the same authors is invoked to forbid alternatives. The benchmark being designed by the authors to target the failure modes E2RAG addresses is a benchmark-design choice, not circular reasoning, because the method's outputs are not derived from the benchmark's labels. Overall, the derivation chain is self-contained in the sense that each reported number comes from running the pipeline, and no step collapses into its own input.

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

The central claim rests on LLM-generated benchmark data, LLM-judged evaluation, and substring-based graph linking. No free parameters are fitted, and no new physical or formal entities are postulated. The main fragility is the evaluation stack: both the benchmark ground truth and the answer scoring come from LLMs, with no human verification or significance testing.

assumptions (4)
  • domain assumption The LLM oracle (GPT-o1-pro, Claude-3.7-Sonnet) generates correct and diverse question-answer pairs for ChronoQA.
    Section 3: the ground truth is the model's retained answer, filtered only by a second LLM pass requiring the excerpt to support the answer; no human verification is reported.
  • domain assumption The LLM judge scores (1-10) are a reliable measure of answer quality.
    Section 5: all rankings rest on mean scores from three judges with no significance testing; Appendix G shows large judge-to-judge scale differences (Claude-3.7-Sonnet 5.36 vs GPT-4.1-mini 9.16 for the top system).
  • domain assumption Surface-form substring matching between an entity name and an event description suffices to link entities to events.
    Eq. (1) and Algorithm 2 build the bipartite edge set B using Name(v) subset of Desc(e); the formal definition omits the same-chunk restriction present in Algorithm 2.
  • domain assumption GPT-4o-mini and text-embedding-3-small are adequate for extraction, embedding, and generation across all pipelines.
    Section 5: all methods share these backbones; if extraction quality is poor for some text genres, comparisons could be distorted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Respecting Temporal-Causal Consistency: Entity-Event Knowledge Graphs for Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/M3SCESE6

@misc{pith2026250605939,
  author       = {Pith},
  title        = {Pith review of: Respecting Temporal-Causal Consistency: Entity-Event Knowledge Graphs for Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3SCESE6}},
  note         = {Machine review of arXiv:2506.05939}
}
read the original abstract

Retrieval-augmented generation (RAG) based on large language models often falters on narrative documents with inherent temporal structures. Standard unstructured RAG methods rely solely on embedding-similarity matching and lack any general mechanism to encode or exploit chronological information, while knowledge graph RAG (KG-RAG) frameworks collapse every mention of an entity into a single node, erasing the evolving context that drives many queries. To formalize this challenge and draw the community's attention, we construct ChronoQA, a robust and discriminative QA benchmark that measures temporal, causal, and character consistency understanding in narrative documents (e.g., novels) under the RAG setting. We then introduce Entity-Event RAG (E^2RAG), a dual-graph framework that keeps separate entity and event subgraphs linked by a bipartite mapping, thereby preserving the temporal and causal facets needed for fine-grained reasoning. Across ChronoQA, our approach outperforms state-of-the-art unstructured and KG-based RAG baselines, with notable gains on causal and character consistency queries. E^2RAG therefore offers a practical path to more context-aware retrieval for tasks that require precise answers grounded in chronological information.

Figures

Figures reproduced from arXiv: 2506.05939 by the authors.

Figure 1
Figure 1. Illustration of two temporally distinct Hermione mentions and their associated events. dow, we do not chunk the text. Instead we use a two-stage pipeline: • Q–A candidates generation: The entire story is provided to the oracle model (GPT-o1-pro (Jaech et al., 2024) and Claude-3.7-Sonnet (Anthropic, 2025)) together with a category list and an instruc￾tion to propose up to diverse question–answer pairs per category (p… view at source ↗
Figure 2
Figure 2. Overall architecture of the E 2RAG framework. in Chapter 9 versus "Miss Granger" in Chapter 16) — distinct and tether it to the event snippet in which it occurs. Intuitively, the structure is as shown in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Heatmap of the scoring for each category and mode. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A snapshot of a part of the Entity-Event KG of [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 8 canonical work pages

  1. [1]

    Chenxin An, Jun Zhang, Ming Zhong, Lei Li, Shansan Gong, Yao Luo, Jingjing Xu, and Lingpeng Kong. 2024. Why does the effective context length of llms fall short? arXiv preprint arXiv:2410.18745

    arXiv 2024

  2. [2]

    Anthropic . 2025. Claude 3.7 sonnet system card. https://www.anthropic.com/claude-3-7-sonnet-system-card. System card, accessed 2025-05-18

    work page 2025

  3. [3]

    Patrice B \'e chard and Orlando Marquez Ayala. 2024. Reducing hallucination in structured outputs via retrieval-augmented generation. arXiv preprint arXiv:2404.08189

    arXiv 2024

  4. [4]

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, pages 2206--2240. PMLR

    2022

  5. [5]

    Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. 2024. Rq-rag: Learning to refine queries for retrieval augmented generation. arXiv preprint arXiv:2404.00610

    arXiv 2024

  6. [6]

    Boyu Chen, Zirui Guo, Zidan Yang, Yuluo Chen, Junze Chen, Zhenghao Liu, Chuan Shi, and Cheng Yang. 2025. Pathrag: Pruning graph-based retrieval augmented generation with relational paths. arXiv preprint arXiv:2502.14902

    arXiv 2025

  7. [7]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130

    arXiv 2024

  8. [8]

    Fiction.live . 2025. https://fiction.live/stories/Fiction-liveBench-Mar-25-2025/oQdzQvKHw8JyXbN87 Fiction.livebench: The first real-world long context benchmark for writers . Accessed: 2025-05-10

    work page 2025

Show all 29 references
  1. [9]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. Precise zero-shot dense retrieval without relevance labels. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762--1777

    2023

  2. [10]

    Muhan Gao, TaiMing Lu, Kuai Yu, Adam Byerly, and Daniel Khashabi. 2024. Insights into llm long-context failures: When transformers know but don’t tell. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7611--7625

    2024

  3. [11]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. https://arxiv.org/abs/2410.05779 Lightrag: Simple and fast retrieval-augmented generation . arXiv preprint arXiv:2410.05779. ArXiv:2410.05779 [cs.IR]

    2024 arXiv

  4. [12]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  5. [13]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    2024 arXiv

  6. [14]

    Ziyan Jiang, Xueguang Ma, and Wenhu Chen. 2024. Longrag: Enhancing retrieval-augmented generation with long-context llms. arXiv preprint arXiv:2406.15319

    2024 arXiv

  7. [15]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information proc...

    2020

  8. [16]

    Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning. arXiv preprint arXiv:2404.02060

    2024 arXiv

  9. [17]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023. Lost in the middle: How language models use long contexts. arXiv preprint arXiv:2307.03172

    2023 arXiv

  10. [18]

    Xiang Liu, Peijie Dong, Xuming Hu, and Xiaowen Chu. 2024. Longgenbench: Long-context generation benchmark. arXiv preprint arXiv:2410.04199

    2024 arXiv

  11. [19]

    Qitan Lv, Jie Wang, Hanzhu Chen, Bin Li, Yongdong Zhang, and Feng Wu. 2024. Coarse-to-fine highlighting: Reducing knowledge hallucination in large language models. arXiv preprint arXiv:2410.15116

    2024 arXiv

  12. [20]

    Xuanfan Ni, Hengyi Cai, Xiaochi Wei, Shuaiqiang Wang, Dawei Yin, and Piji Li. 2024. Xl ^2 bench: A benchmark for extremely long context understanding with long-range dependencies. arXiv preprint arXiv:2404.05446

    2024 arXiv

  13. [21]

    OpenAI. 2024. text-embedding-3-small : Openai embedding model. https://platform.openai.com/docs/models/text-embedding-3-small. Model release announced 25 Jan 2024

    2024

  14. [22]

    OpenAI. 2025. Introducing gpt–4.1 in the api. https://openai.com/index/gpt-4-1. Accessed: 20 May 2025

    2025

  15. [23]

    Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921

    2024 arXiv

  16. [24]

    Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. arXiv preprint arXiv:2104.07567

    2021 arXiv

  17. [25]

    Chandana Sree Mala, Gizem Gezici, and Fosca Giannotti. 2025. Hybrid retrieval for hallucination mitigation in large language models: A comparative analysis. arXiv e-prints, pages arXiv--2504

    2025

  18. [26]

    Minzheng Wang, Longze Chen, Cheng Fu, Shengyi Liao, Xinghua Zhang, Bingli Wu, Haiyang Yu, Nan Xu, Lei Zhang, Run Luo, et al. 2024. Leave no document behind: Benchmarking long-context llms with extended multi-doc qa. arXiv preprint arXiv:2406.17419

    2024 arXiv

  19. [27]

    Shu Wang, Yixiang Fang, Yingli Zhou, Xilin Liu, and Yuchi Ma. 2025. Archrag: Attributed community-based hierarchical retrieval-augmented generation. arXiv preprint arXiv:2502.09891

    2025 arXiv

  20. [28]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  21. [29]

    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.