Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Simple Mechanistic Explanations for Out-Of-Context Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Out-of-context reasoning in fine-tuned language models often reduces to a single steering vector.

desk verdict A real steering-vector finding is overgeneralized in the abstract; worth review after a scope fix. read the letter →

arxiv 2507.08218 v2 pith:UWRLGJ4U submitted 2025-07-10 cs.CL cs.LG

Atticus Wang , Joshua Engels , Oliver Clive-Griffin , Senthooran Rajamanoharan , Neel Nanda This is my paper
classification cs.CLcs.LG
keywords out-of-contextreasoningsteeringvectorsLoRAmechanisticinterpretabilityfine-tuningactivationengineeringmodeldiffinggeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that out-of-context reasoning (OOCR) in fine-tuned language models often arises from a simple mechanism: the LoRA adapter effectively adds a single near-constant steering vector to the model's activations, shifting the model toward a general concept. The authors support this by showing that LoRA output differences are nearly collinear across tokens and that steering vectors extracted from LoRA, or trained directly, reproduce OOCR on risky/safe behavior, risk backdoors, city identification, and function description tasks. They also show that a backdoor task that looks conditional can be implemented by an unconditional steering vector. If this is right, a surprising form of out-of-distribution generalization can be understood and controlled as a single learned direction.

What carries the argument

The central object is the 'LoRA output difference vector': for a given token, the difference between the MLP output with and without the LoRA adapter. When these vectors are almost collinear across tokens, the adapter is equivalent to adding one fixed 'steering vector' to the residual stream. The steering vector is a fixed vector added to model activations at inference time to shift behavior; the paper shows it can be extracted from the LoRA update or trained directly, and that it points the model toward a general concept representation already present in the base model.

What would settle it

Find an OOCR task from the original literature that is reproduced with full fine-tuning or all-layer LoRA, and check whether the fine-tuned behavior can be reproduced by adding a single constant vector to the base model; if no single steering vector reproduces it, the paper's representativeness assumption and the broad explanation fail for that task. Alternatively, measure LoRA output differences on a large, diverse corpus of prompts for a task where single-layer LoRA shows OOCR; if the top principal component explains far less variance than the near-collinearity reported here, the mechanism is not a single steering vector.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that fine-tuning via a single-layer LoRA adapter on an MLP block learns a single output direction: the per-token differences between the adapter's output and the base output have pairwise absolute cosine similarities close to one, so the adapter is effectively adding the same vector at every token. Extracting this direction (by first principal component or by unitizing and averaging the differences) and adding it to the base model as a steering vector yields out-of-context reasoning at levels comparable to the full LoRA adapter. Directly training a steering vector from scratch also induces OOCR on the same tasks, and the learned vectors are not the naive 'codename-to-real-name' directions. Even the risk-backdoor task, which appears to require conditional behavior, is reproduced by an unconditional steering vector added to the final token, which the authors attribute to the jailbreak token's value vectors aligning with the steering direction.

Load-bearing premise

The claim that many OOCR instances are steering vectors rests on single-layer LoRA being representative of the fine-tuning setups where OOCR was originally demonstrated, since all-layer LoRA fails on the Functions and Locations tasks and the collinearity measurement uses only two passages at one layer.

Editorial extensions

If this is right

  • If OOCR is often a single steering vector, then trained steering vectors can reproduce OOCR without fine-tuning, providing a cheap and interpretable method for inducing, controlling, or auditing the behavior.
  • The steering-vector account predicts that the success of OOCR depends on the base model already containing a general concept representation, so tasks with no such representation should not show OOCR under LoRA fine-tuning.
  • For backdoor-like behaviors, an unconditional steering vector can implement a conditional trigger, so the presence of a conditional behavior does not imply a conditional mechanism.
  • The logit-lens interpretability of the natural steering vectors offers a token-level vocabulary for describing what concept the fine-tuned model has moved toward.
  • Because the mechanism is a single learned direction, understanding when OOCR appears reduces to understanding when gradient descent finds such a direction, linking OOCR to the study of simplicity biases in SGD.

Reading between the lines

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

  • We infer a practical diagnostic: for any new OOCR task, one could compute the top principal component of the LoRA output differences as a fast proxy for whether the task will admit the steering-vector explanation, without training or interpreting full circuits.
  • We infer that the failure of all-layer LoRA on Functions and Locations hints that multi-layer adapters can 'wash out' the single-direction update, so the steering-vector mechanism may be specific to low-rank, single-module fine-tuning rather than to OOCR in general.
  • We infer that the backdoor result extends to safety more broadly: an apparent conditional capability (acting only under a trigger) can be produced by an unconditional additive change, so interpretability methods that search for conditional circuitry may miss the actual mechanism.
  • We infer that the near-collinearity measurement, based on two passages and the final 20 tokens, suggests a testable extension: replicating on a much larger set of diverse contexts to see whether the single-direction account survives.
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 mechanistic explanation for out-of-context reasoning (OOCR): LoRA fine-tuning effectively adds a near-constant steering vector to model activations, shifting the model toward a general concept, and this shift produces the surprising OOCR generalization. The authors study four tasks (Risky/Safe, Risk Backdoor, Locations, Functions) using Gemma 3 12B with rank-64 LoRA on MLP layers. They report that single-layer LoRA adapters produce output differences that are nearly collinear, that a 'natural' steering vector extracted from LoRA reproduces part of the OOCR behavior, and that directly trained steering vectors can also induce OOCR. They additionally report that an unconditional steering vector can implement an in-distribution conditional backdoor behavior, though OOCR test performance on the backdoor task is low.

Significance. If the broad claim held, this would be a valuable simplification: OOCR could be understood as a single-direction activation shift rather than a complex set of circuit changes. The paper's concrete strengths are the public code and datasets, the direct training of steering vectors from scratch as an independent confirmation of the mechanism, the comparison with a naive concept-vector baseline, and the honest reporting of several failures. The contribution is real but narrower than the abstract claims: the steering-vector mechanism is demonstrated in single-layer LoRA at selected layers, while all-layer LoRA fails on two of the four main tasks, and the backdoor result is only in-distribution. The paper is therefore a suggestive case study rather than a demonstrated general explanation of OOCR in the literature.

major comments (4)
  1. [§4.1, Figure 2; Abstract] The abstract claims that 'many instances of OOCR in the literature' have a constant-steering-vector explanation, but the evidence in Figure 2 shows that all-layer LoRA, the configuration closest to the standard fine-tuning setup in the original OOCR reports, achieves negligible OOCR accuracy on the Functions and Locations tasks. The steering-vector result is demonstrated only for single-layer LoRA at hand-selected layers (layer 22 for Risky/Safe, layer 2 for Functions, layer 15 for Locations). No experiment transfers the single-layer finding to the all-layer or full-fine-tuning setting in which OOCR was originally observed. This is the load-bearing step for the headline claim and needs either direct transfer evidence or a substantially weakened claim.
  2. [§4.2, Figure 4] The near-collinearity claim is based on only two passages (one out-of-distribution passage and the first training example) and the last 20 tokens, evaluated at the best-performing layer per task. These 40 vectors are not sufficient to establish that the LoRA adapter behaves as a constant steering vector across contexts, especially because the same test set was used to select the layer in Section 4.1. The authors should either report collinearity distributions over many more contexts and token positions without test-set layer selection, or explicitly constrain the claim to the measured contexts.
  3. [§4.1, Figure 3] The best-performing layer for each task appears to be selected using the OOCR test set, and the reported accuracies (e.g., the single-layer bars in Figure 2) are then computed on that same test set. This test-set selection biases the comparison in favor of the single-layer steering-vector explanation. A fair comparison requires selecting layers on a validation split, or reporting the full layer sweep without choosing the maximum, and then evaluating on a held-out test set.
  4. [§5.4, Figure 2] The backdoor result is presented in the abstract as evidence that the steering-vector explanation holds 'even for a task that seems like it must involve conditional behavior,' but Figure 2 shows that both LoRA and steering vectors achieve low OOCR test accuracy on the backdoor datasets, and Section 5.4 reports high in-distribution validation accuracy only. The paper's own caveat admits that this does not show whether steering vectors can learn OOCR backdoors. The abstract and conclusions should be revised to reflect that the backdoor finding is an in-distribution phenomenon, not an instance of OOCR.
minor comments (5)
  1. [Appendix A.1, Figures 2 and 3] Appendix A.1 states that error bars are over 5 runs with different seeds, but the main bar charts (Figures 2 and 3) do not display visible error bars; please add them or state explicitly where they are shown.
  2. [§5.4] In the sentence beginning 'an steering vector added to a fixed token position,' the article should be 'a steering vector.'
  3. [Contributions] The Contributions section credits 'SE provided regular ideas and feedback,' but the author list uses the initials SR for Senthooran Rajamanoharan; please correct the abbreviation.
  4. [Figure 2] The caption says 'The single layer bars use layer 22,' but the other single-layer results in the same figure use layers 2 and 15; the caption should specify the layer used for each dataset.
  5. [§4.2] The phrase 'we examine the vector that the LoRA matrix adds to the model (this is equivalent to the difference between the output of the model component with and without the LoRA adapter)' is clear, but the following sentence about pairwise cosine similarities is grammatically tangled and should be rephrased.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation collapses into its inputs; the steering-vector mechanism is supported by held-out tests and direct training, but the per-task 'best layer' is selected on the same test data used to report OOCR accuracy, giving a mild selection-on-the-evaluation-target effect.

  1. fitted input called prediction [Section 4.1, paragraph beginning 'We next look at just the best-performing layer for each task', and Figure 3 caption 'Test set performance'.]
    "We next look at just the best-performing layer for each task: layer 22 for Risky/Safe behavior, layer 2 for the function f (x) = 3x + 2, and layer 15 for the location Tokyo."

    The 'best-performing layer' is selected by maximizing the same per-task OOCR test accuracy that is then reported as the central evidence in Figure 2, Figure 3, and Figure 5, and it is also the layer at which natural steering vectors are extracted and evaluated. In other words, a discrete hyperparameter (the layer index) is fitted to the evaluation target, and the resulting test-set accuracy is subsequently presented as the performance of single-layer LoRA and of the steering vectors at that layer.

full rationale

The paper's central derivation is not circular in the strong sense. The near-collinearity claim in Section 4.2 is an empirical measurement of LoRA output differences on an unrelated OOD passage and the first training example; the natural steering vectors are extracted from these LoRA differences, not from the OOCR test set, and testing them on held-out tasks is a genuine generalization check. Directly training steering vectors from scratch in Section 5.1 provides an independent route to OOCR, and the comparison against the 'naive steering vector' in Section 5.2 shows the learned vectors are not trivially predetermined by the task setup. The backdoor section is transparent about the failure to achieve OOCR test accuracy, and the causal patch in Figure 9 is separate evidence. The main circularity-adjacent issue is the selection of the per-task 'best layer' using test-set accuracy, which means the headline comparisons at that layer are partly self-selected; this is a methodological leakage rather than a definitional reduction. The abstract's 'many instances of OOCR in the literature' also overgeneralizes from single-layer LoRA, since the paper itself reports that all-layer LoRA fails on Functions and Locations; that is a correctness and scope concern, not a circularity. No load-bearing self-citation or imported uniqueness theorem is used, and the core mechanism claim stands on independent evidence, so the score is a mild 2 rather than a higher circularity score.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or theoretical entities. The steering vector is an existing concept, though the token-conditional variant used for Functions/Locations is a task-specific instantiation rather than an invented entity. The main assumptions are domain-level: that the base model's representations are steerable, that the benchmarks capture OOCR, and that the analysis methods (LoRA difference vectors, activation patching, logit lens) are valid.

free parameters (4)
  • LoRA rank, dropout, alpha = rank=64, dropout=0.05, alpha=32
    Hyperparameters chosen by hand for all LoRA experiments; they affect the LoRA output difference vectors and the steering vector equivalence.
  • Best-performing layer per task = layer 22 (Risky/Safe), layer 2 (Functions), layer 15 (Locations)
    The paper selects the layer with highest OOCR test accuracy for each task before analyzing LoRA directions, introducing a post-hoc selection step.
  • Number of tokens used for LoRA difference analysis = 20 (last tokens)
    The near-collinearity claim is based on the last 20 tokens of two passages; early tokens are excluded to avoid spikes, which is a manual choice.
  • Steering vector norm initialization = 1
    Directly trained steering vectors are initialized with norm 1, but the trained norm is a fitted quantity; the initialization is a hyperparameter.
assumptions (5)
  • domain assumption The base model contains latent representations of the target concepts (risk, safety, city, function) that can be activated by a steering vector.
    This is the core of the hypothesis in Section 5.1 and is assumed throughout; no independent evidence is given that these concepts are localizable as a single direction.
  • domain assumption The OOCR test sets (self-report of risk, city identification, function description) are valid measures of out-of-context reasoning as defined by Berglund et al.
    The paper imports these benchmarks from prior work without re-validating them for the Gemma 3 model or the single-layer LoRA setup.
  • domain assumption The LoRA output difference vector, computed as the difference between the component's output with and without LoRA, accurately captures the learned change.
    The collinearity analysis in Section 4.2 relies on this difference as the unit of analysis; it ignores interactions with other layers and attention.
  • domain assumption Activation patching in Section 5.4 is a valid causal intervention for identifying the role of QK attention patterns on the last token.
    The backdoor explanation uses patching to infer causal involvement, but the method relies on standard interpretability assumptions about linearity and locality.
  • domain assumption The logit lens reliably reveals the semantic content of a steering direction.
    The interpretability results in Sections 4.4 and 5.3 take top logits as evidence for risk/safety meaning, which is a known but imperfect probe.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simple Mechanistic Explanations for Out-Of-Context Reasoning." pith.science (2026). https://pith.science/paper/UWRLGJ4U

@misc{pith2026250708218,
  author       = {Pith},
  title        = {Pith review of: Simple Mechanistic Explanations for Out-Of-Context Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWRLGJ4U}},
  note         = {Machine review of arXiv:2507.08218}
}
read the original abstract

Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and in many other concept-related domains, causing the surprising generalization. Moreover, we can directly train steering vectors for these tasks from scratch, which also induces OOCR. We find that our results hold even for a task that seems like it must involve conditional behavior (model backdoors); it turns out that unconditionally adding a steering vector is sufficient. Overall, our work presents one explanation of what gets learned during fine-tuning for OOCR tasks, contributing to the key question of why LLMs can reason out of context, an advanced capability that is highly relevant to their safe and reliable deployment.

Figures

Figures reproduced from arXiv: 2507.08218 by the authors.

Figure 1
Figure 1. A figure containing our main results. Bottom: Adding steering vectors to an LLM can explain OOCR, see [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Test accuracies for different OOCR tasks; the test datasets measure whether the model can implicitly use information from fine-tuning. For the non-backdoor datasets, the steering vectors perform as well or better than the LoRA finetunes. The single layer bars use layer 22. The backdoors are averaged over the four backdoor system prompts. Although we get high validation accuracy on the backdoor tasks (see Section 5.4… view at source ↗
Figure 4
Figure 4. Histogram of cross-token pairwise cosine similarities of LoRA output difference vectors of different tasks, after taking the absolute value. Note that the histograms for each task also include pairwise cosine similarities between the LoRA difference vectors of in and out of distribution prompts. only on MLP blocks (not attention). We investigate training LoRA in two different ways: 1. All Layers: We train an adapter… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: OOCR test accuracies for the Functions task with the function f(x) = 3x + 2. In early layers, learning a LoRA adapter and learning a conditional steering vector both enable OOCR. Two types of “natural” steering vectors (taking the first principal component of LoRA diff…
Figure 7
Figure 7. Figure 7: show that many layers’ natural steering vectors are indeed interpretable with the logit lens. Roughly, the percentage of interpretable tokens goes down following the accuracy and logit difference drop from [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: (a) For a fixed set of functions in the Functions task, the cosine similarity between their na¨ıve steering vectors and their learned steering vectors. In particular learned vectors do not have bigger cosine similarity with their corresponding na¨ıve vectors. (b) For a…
Figure 10
Figure 10. Figure 10: OOCR test accuracies for the Locations task with the city Tokyo. In layers 14–16, the LoRA adapters and steering vectors achieves non-trivial OOCR generalization. A.3. Out-of-distribution passage used to analyze LoRA outputs One of the more famous episodes of this sor…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Revealing Hidden Model Behaviors with Task-Specific Self-Reports

    cs.CL 2026-07 conditional novelty 6.0 of 10

    SAR detects every implanted hidden behavior across eight Qwen3-14B settings and halves IA’s hallucination rate by aligning self-report activations to a contrastive behavior direction under a coherent-English stabilizing cap.

  2. Revealing Hidden Model Behaviors with Task-Specific Self-Reports

    cs.CL 2026-07 conditional novelty 6.0 of 10

    A per-model LoRA adapter trained on a fine-tuned model's own data can get it to state its hidden behavior in plain English across seven tested behaviors.

Reference graph

Works this paper leans on

19 extracted references · 6 canonical work pages · cited by 1 Pith paper

  1. [1]

    C., Balesni, M., Kaufmann, M., Tong, M., Korbak, T., Kokotajlo, D., and Evans, O

    Berglund, L., Stickland, A. C., Balesni, M., Kaufmann, M., Tong, M., Korbak, T., Kokotajlo, D., and Evans, O. Taken out of context: On measuring situational awareness in llms. arXiv preprint arXiv:2309.00667, 2023

    arXiv 2023

  2. [2]

    Tell me about yourself: Llms are aware of their learned behaviors

    Betley, J., Bao, X., Soto, M., Sztyber-Betley, A., Chua, J., and Evans, O. Tell me about yourself: Llms are aware of their learned behaviors. arXiv preprint arXiv:2501.11120, 2025

    arXiv 2025

  3. [3]

    and Cohan, A

    Dunefsky, J. and Cohan, A. Investigating generalization of one-shot llm steering vectors, 2025. URL https://arxiv.org/abs/2502.18862

    arXiv 2025

  4. [4]

    Alignment faking in large language models

    Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDiarmid, M., Marks, S., Treutlein, J., Belonax, T., Chen, J., Duvenaud, D., et al. Alignment faking in large language models. arXiv preprint arXiv:2412.14093, 2024

    arXiv 2024

  5. [5]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

    2022

  6. [6]

    Chris olah's views on AGI safety

    Hubinger, E. Chris olah's views on AGI safety. https://www.alignmentforum.org/posts/X2i9dQQK3gETCyqh2/chris-olah-s-views-on-agi-safety, November 2019. Accessed: 2025-05-19

    work page 2019

  7. [7]

    Gemma 3 technical report

    Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S., Matejovicova, T., Ram \'e , A., Rivi \`e re, M., et al. Gemma 3 technical report. arXiv preprint arXiv:2503.19786, 2025

    arXiv 2025

  8. [8]

    Gemini 2.5: Our most intelligent ai model

    Kavukcuoglu, K. Gemini 2.5: Our most intelligent ai model. ://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/, March 2025. Accessed June 10, 2025

    work page 2025

Show all 19 references
  1. [9]

    Sparse crosscoders for cross-layer features and model diffing

    Lindsey, J., Templeton, A., Marcus, J., Conerly, T., Batson, J., and Olah, C. Sparse crosscoders for cross-layer features and model diffing. Transformer Circuits Thread, 2024. URL https://transformer-circuits.pub/2024/crosscoders/index.html. Accessed: YYYY-MM-DD

    2024

  2. [10]

    Auditing language models for hidden objectives

    Marks, S., Treutlein, J., Bricken, T., Lindsey, J., Marcus, J., Mishra-Sharma, S., Ziegler, D., Ameisen, E., Batson, J., Belonax, T., et al. Auditing language models for hidden objectives. arXiv preprint arXiv:2503.10965, 2025

    2025 arXiv

  3. [11]

    Robustly identifying concepts introduced during chat fine-tuning using crosscoders

    Minder, J., Dumas, C., Juang, C., Chugtai, B., and Nanda, N. Robustly identifying concepts introduced during chat fine-tuning using crosscoders. arXiv preprint arXiv:2504.02922, 2025

    2025

  4. [12]

    nostalgebraist. 2020. URL https://www.lesswrong. com/posts/AcKRB8wDpdaN6v6ru/ interpreting-gpt-the-logit-lens

    2020

  5. [13]

    R., Haklay, T., Belinkov, Y., and Bau, D

    Prakash, N., Shaham, T. R., Haklay, T., Belinkov, Y., and Bau, D. Fine-tuning enhances existing mechanisms: A case study on entity tracking. arXiv preprint arXiv:2402.14811, 2024

    2024 arXiv

  6. [14]

    M., Ilyas, A., and Madry, A

    Shah, H., Park, S. M., Ilyas, A., and Madry, A. Modeldiff: A framework for comparing learning algorithms. In International Conference on Machine Learning, pp.\ 30646--30688. PMLR, 2023

    2023

  7. [15]

    Subramani, N., Suresh, N., and Peters, M. E. Extracting latent steering vectors from pretrained language models. arXiv preprint arXiv:2205.05124, 2022

    2022 arXiv

  8. [16]

    Analysing the generalisation and reliability of steering vectors

    Tan, D., Chanin, D., Lynch, A., Paige, B., Kanoulas, D., Garriga-Alonso, A., and Kirk, R. Analysing the generalisation and reliability of steering vectors. Advances in Neural Information Processing Systems, 37: 0 139179--139212, 2024

    2024

  9. [17]

    B., and Evans, O

    Treutlein, J., Choi, D., Betley, J., Marks, S., Anil, C., Grosse, R. B., and Evans, O. Connecting the dots: Llms can infer and verbalize latent structure from disparate training data. Advances in Neural Information Processing Systems, 37: 0 140667--140730, 2024

    2024

  10. [18]

    M., Thiergart, L., Leech, G., Udell, D., Vazquez, J

    Turner, A. M., Thiergart, L., Leech, G., Udell, D., Vazquez, J. J., Mini, U., and MacDiarmid, M. Steering language models with activation engineering. arXiv preprint arXiv:2308.10248, 2023

    2023 arXiv

  11. [19]

    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 6, 2026 · model on record in the stance chip above.