Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear author,
Thank you very much for the opportunity to review your article, both the topic and the approach are very innovative.
LLMs themselves are currently in full development and any study on their evolution is more than welcome. However, care must be taken with the strength of the statements to avoid falling into speculation. Below are some recommendations for a slight improvement of your article:
- lines 81-82 announce for Section 3 testable examples of LLM failure "that directly trace to the absence of counterhactual capacity" and in section 3 (lines 128-157) are mentioned very briefly in fact 3 types of LLM failures. Attention, in line 131 the presentation of 4 categories of failures is announced (and only 3 are presented). An expansion of this section, with more references and data from concrete studies would bring more consistency to the article.
- lines 254-255: "LLMs trained on text alone learn the expression of causal reasoning, not its structure (Gopnik, 2009)" - the LLMs are dated somewhere after 2017, the reference is not really appropriate for the idea of ​​an LLM
- lines 263-267 are you sure that a pattern match system cannot answer the question "how did you arrive at this conclusion?" Because if a reasoning is copied, its argumentation can also be copied. Moreover, the LLMs present the sources used when providing an answer, I would recommend a slight reformulation of this paragraph
- line 272 typo "nost likely"

Once again, congratulations, your article is very interesting!

Author Response

RESPONSE TO REVIEW 1

Review 1 was positive overall and focused on three main areas: (1) expansion of Section 3 with evidence, (2) clarification of the Gopnik citation, and (3) qualification of pattern-matching claims. We address each below.

Issue 1.1: Section 3 Expansion and Missing Fourth Category

Reviewer 1 noted: "Section 3 promises FOUR categories of failure but presents only THREE. The three subsections are: Adversarial fragility, Collider reasoning, Individual counterfactuals. The reviewer requested expansion with more references and actual research evidence."

Revision:
We have (a) renamed Section 3 to reflect the proper scope ("Where Text-Only LLM Reasoners Fail"); (b) expanded it from three to FOUR failure categories by adding a fourth: "Counterfactual Invariance Under Causally Equivalent Reformulations" [Section 3.4]. This new category provides a concrete diagnostic criterion for distinguishing fluent language about causality from stable causal reasoning.

Additionally, we have: (c) qualified each failure mode to clarify its diagnostic scope; (d) added the meta-commentary that "These failure modes are not individually diagnostic; their relevance is cumulative" [Section 3, opening]; and (e) added specific citations to recent empirical work: Jin et al. 2024 (Corr2Cause benchmark), Joshi et al. 2024 (LLM causal fallacies), Zou et al. 2023 (transferable adversarial attacks).

 

Issue 1.2: Gopnik (2009) Citation - Clarification

Reviewer 1 noted that the paper states "LLMs trained on text alone learn the expression of causal reasoning, not its structure (Gopnik, 2009)," but Gopnik 2009 is about child development, not AI systems.

Revision:
The citation context has been clarified. The sentence now reads: "By contrast, text-only LLMs are trained primarily on linguistic traces of such reasoning. For this reason, their causal competence should not be inferred from fluent causal language alone" [Section 6]. The Gopnik reference supports the developmental principle (that causal cognition emerges through active exploration, not passive linguistic exposure), not the claim about LLMs specifically. The connection is now made explicit.

 

Issue 1.3: Pattern-Matching Claims - Qualification

Reviewer 1 noted that the statement "pattern-matching systems cannot answer 'How did you arrive at this conclusion?'" may be too strong, since LLMs do provide reasoning chains.

Revision:
Section 3 and the implications section have been revised to distinguish between "fluent explanation" and "causal model." The statement is now: "Systems may produce fluent explanations without possessing a causal model. The relevant test is whether their explanations remain stable under causally irrelevant reformulations and change appropriately under selective causal interventions" [Section 6]. This avoids claiming that LLMs cannot produce explanations; it instead specifies the criterion (stability under reformulation) by which causal reasoning can be distinguished from fluent pattern-matching.

 

Issue 1.4: Typo (nost → most)

Reviewer 1 flagged "nost likely" (line 272).

Revision: Fixed. The sentence now reads: "...most likely not through formal probabilistic calculus..." [Conclusion].

Reviewer 2 Report

Comments and Suggestions for Authors

This paper offers a compelling and timely argument, effectively synthesizing cognitive science and neuroscience to critique current AI's lack of counterfactual reasoning. Its core proposal—that AI should seek to replicate the brain's heuristic approximations of causality rather than encode formal causal calculus—is a distinctive and valuable contribution.

To strengthen the manuscript, please address the following points:

1. Clarify the Mechanism of "Simulation".
The terms "simulating counterfactually" and "simulation-based" are central yet under-defined. Please add a brief explanation clarifying how Johnson-Laird's "mental models" theory implements cognitive simulation (e.g., via token-based scenario construction). Contrast this with the requirements for an AI "causal world model" (e.g., a learnable, manipulable representation of structural equations). This will make the metaphor more operational for computational research.

2. Provide Concrete Research Examples.
Sections 5 and 6 outline a promising NeuroAI blueprint but remain high-level. To enhance practical guidance, please add one or two specific research pathway examples. For instance: (a) Designing neural modules inspired by the Default Mode Network's generative dynamics for internal counterfactual exploration. (b) Using in vitro neural cultures to study causal sequence learning and translating the observed dynamics into AI algorithms.

3. Correct Editorial Oversights.
A thorough proofreading is needed. Key issues include: citation mismatches (e.g., Goodfellow et al. 2014 text vs. 2015 in reference list; Pearl & Mackenzie text: 2020, list: 2018), the typo "reasons casually" (should be "causally").

Author Response

RESPONSE TO REVIEW 2

Review 2 was favorable but identified three areas for strengthening: (1) clarification of simulation mechanisms, (2) concrete research pathways, and (3) editorial corrections. We address each.

Issue 2.1: Mechanism of Simulation - Johnson-Laird Clarification

Reviewer 2 noted: "The terms 'simulating counterfactually' and 'simulation-based' are central to the argument but under-defined. Please explain how Johnson-Laird's mental models theory implements cognitive simulation, and contrast this with Pearl's formal causal world model."

Revision:
We have substantially expanded Section 4.2 ("The Role of Mental Models") to provide exactly this clarification. The revised section now explains:

(a) What a mental model is: "A mental model can be understood as a token-based representation of a scenario: a temporary configuration of entities, relations, events, and constraints that stands for a possible situation" [Section 4.2, revised paragraph 1].

(b) How mental models implement counterfactual reasoning: by constructing two tokens (actual and hypothetical), preserving identity while selectively modifying one element, and evaluating consequences. [Section 4.2, revised paragraph 2].

(c) Explicit contrast with Pearl's framework: The text now states: "This differs from Pearl's formal causal model. In Pearl's framework, counterfactual inference requires a structural causal model: a set of variables connected by structural equations... Mental models are less formal but cognitively plausible approximations of this requirement..." [Section 4.2, new paragraph 3].

This directly addresses the reviewer's request to make the distinction operational for computational research. The section concludes by emphasizing the central point: the brain need not implement formal causal calculus to approximate its computational functions.

 

Issue 2.2: Concrete Research Pathways

Reviewer 2 noted: "Sections 5 and 6 outline a NeuroAI blueprint but remain largely high-level and programmatic. Please add one or two concrete research pathway examples (e.g., DMN-inspired modules in transformers, MEA cultures for causal learning)."

Revision:
We have entirely restructured Section 5 and added a new subsection 5.1 titled "Two Concrete NeuroAI Research Pathways" [NEW, Section 5.1]. This subsection contains two detailed pathways:

Pathway 1: "DMN-inspired generative modules for counterfactual exploration in transformers" [Section 5.1.1]. This pathway explains how the DMN principle (scenario construction with reduced sensory constraint) could inform a latent scenario generator in a transformer, preserving entity identity while selectively modifying causal variables. Success criteria are specified: invariance across paraphrases, observation/intervention distinction, stable individual-level inference.

Pathway 2: "In vitro neural cultures and causal sequence learning on multi-electrode arrays" [Section 5.1.2]. This pathway proposes using MEA cultures to study how recurrent networks learn temporal contingencies, respond to perturbation, and update models—producing AI-relevant principles for sequence learning and model revision.

Both pathways specify: (1) the biological principle, (2) the AI implementation, and (3) how success would be evaluated. This makes the NeuroAI program operationally concrete, addressing the reviewer's core concern.

 

Issue 2.3: Editorial Corrections and Reference Dates

Reviewer 2 flagged date mismatches in references.

Revision: All reference dates have been corrected and verified. The key corrections were:
• Goodfellow et al.: Now correctly listed as 2015 (in-text and reference list).
• Hubel & Wiesel: Now correctly listed as 1959 (primary neurophysiology reference).
• Marr: Now correctly listed as 1982 (Vision).
• Pearl & Mackenzie: Now correctly listed as 2020.
• Kahneman: Updated in-text citations to 2013 edition of Thinking, Fast and Slow.

DOI links have been added where available.

Reviewer 3 Report

Comments and Suggestions for Authors

<General Summary>

The study argue that current LLMs remain largely confined to association, that counterfactual reasoning is the key missing faculty separating statistical AI from human cognition, and that NeuroAI should study how brains approximate counterfactual reasoning rather than directly implement Pearlian formalism. The current version substantially overstates several claims, treats heterogeneous literatures too selectively, and does not yet provide the "concrete, domain-specific examples" of LLM failure that it promises.  The manuscript also contains reference issues, including missing citations, mismatched dates, apparently uncited references, and at least one irrelevant author self-citation which weaken the scientific reliability and publication readiness.

<Major Issues>

The study's novelty and conceptual advance are not yet sufficiently established. The core idea that current LLMs often lack robust causal and counterfactual competence is important, but it is not new as framed; the manuscript itself relies heavily on Pearl, Pearl and Mackenzie, Zecevic et al., Jin et al., and broader NeuroAI arguments. The paper needs to state more precisely what it adds beyond existing claims that LLMs "talk causality" without possessing causal models, and beyond current benchmark work showing weak out-of-distribution causal inference. Jin et al.'s Corr2Cause benchmark, for example, already directly tests whether LLMs can infer causation from correlational statements and reports weak generalization beyond familiar distributions, while Zecevic et al. explicitly argue that LLMs may recite causal knowledge without being causal systems. A publishable Viewpoint would need a sharper original synthesis, a new taxonomy, a clearer research agenda, or a more rigorous integration of causal inference, cognitive development, neuroscience, and AI architecture.

The central claim that "all current machine learning systems operate" at association level is too broad and should be substantially narrowed. The manuscript moves between "LLMs", "current AI systems", "statistical AI", and "all current machine learning systems" as though these terms were interchangeable, but they are not. Text-only LLMs, multimodal foundation models, model-based reinforcement learning agents, causal discovery systems, robotic systems, world-model approaches, and hybrid neuro-symbolic systems raise different questions. The paper can defensibly argue that many current LLMs do not robustly support explicit individual-level counterfactual reasoning, but it cannot simply equate all AI with Level 1 association. The authors should define the target class of systems, specify the behavioral criterion for Level 2 versus Level 3 competence, and address alternative explanations for observed failures, including benchmark design, insufficient grounding, missing tools, poor prompting, distribution shift, uncertainty calibration, and lack of explicit causal-variable representation.

The section on LLM failures is currently underdeveloped and internally inconsistent. The text states that it will provide "four categories of failure", but the manuscript presents only three subsections: adversarial fragility, collider reasoning, and individual counterfactuals. More importantly, these examples are asserted rather than demonstrated. Adversarial fragility is not, by itself, diagnostic of absent counterfactual reasoning; Wallace et al.'s work on universal adversarial triggers shows severe vulnerability and dataset/model biases in NLP systems, but such vulnerability could arise from many mechanisms besides a missing Level 3 causal model. Collider reasoning is closer to Level 2 intervention/conditioning than Level 3 counterfactual reasoning, so the paper should not use it as evidence for the same deficit without clearer logic. For individual counterfactuals, the paper needs actual task examples, model outputs, controls, and failure analyses, or else a careful review of existing experimental studies. Without such evidence, the strongest empirical claims read as plausible commentary rather than supported scientific argument.

The developmental and neuroscientific sections overinterpret relevant but not decisive evidence. Meltzoff's 1995 work is more accurately described as evidence that infants can infer and re-enact intended actions, not direct evidence that 18-month-olds construct Pearl-style counterfactual models. Similarly, DMN, hippocampal, and prefrontal findings support episodic simulation, scene construction, future thinking, regret, and option-value comparison, but these processes are not identical to causal counterfactual inference. Hassabis et al. tested whether patients with bilateral hippocampal damage could construct new imagined experiences, which supports a role for the hippocampus in imagination and scene construction but does not alone establish a neural substrate for formal counterfactual causal reasoning. The manuscript should separate mental simulation, episodic construction, regret/fictive-error processing, causal learning, intervention, and counterfactual inference. It should also soften unsupported evolutionary claims, especially statements implying that causal cognition has been documented across roughly 500 million years or that mammalian brain evolution directly maps onto the missing faculty in LLMs.

The NeuroAI argument is promising but historically oversimplified. The manuscript states that convolutional networks, attention, and reinforcement learning emerged from neuroscience rather than formal theory, but this is only partly true. Hubel and Wiesel's visual cortex work is relevant to the history of convolutional inspiration, but reinforcement learning has deep roots in control theory, dynamic programming, and decision theory as well as later neuroscientific links to dopamine prediction errors. Similarly, transformer self-attention was not derived from prefrontal-parietal attention mechanisms in the way the text implies. Zador et al. do argue for investing in NeuroAI and embodied evaluation, but that does not justify replacing formal causal modeling with vague biological inspiration. The manuscript should make the proposed research program concrete: which neural computations should be studied, which species or preparations are appropriate, which behavioral tasks isolate Level 2 and Level 3 cognition, which architectural mechanisms should be tested, and how success should be evaluated against causal benchmarks.

The reference integrity requires major correction. The manuscript cites several works in the text that are absent from the reference list, including Dusi 2026, Zhang et al. 2025, Wallace et al. 2019, Marr 1996, Miller and Cohen 2001, Coricelli et al. 2005, Padoa-Schioppa and Assad 2006, and Johnson-Laird 1995. It also cites Pearl and Mackenzie 2020 in the text while listing the 2018 book, cites Goodfellow et al. 2014 in text while listing 2015, cites Hubel and Wiesel 1977 in text while listing the 1959 receptive-field paper, and includes Grimaldi et al. 2023 in the references despite no clear relevance or in-text use. The Goodfellow adversarial-examples paper is available as a 2014 arXiv preprint and commonly associated with ICLR 2015, so the citation format should be internally consistent. These affect whether the manuscript accurately represents the literature.

<Minor Issues>

The manuscript needs careful proofreading and tighter terminology. Errors include "nost likely", instead of "most likely"; "casually", instead of "causally"; "formal laws theory", "calculus(Gopnik", and "neuroscience.:". The manuscript also uses incomplete reference notation, such as "s.d.". The 2025 article header also conflicts with the February 2026 interview in the introduction, so the author should check the date or explain the timing. Key terms, including "counterfactual reasoning", "causal reasoning", "causal cognition", "causal world model", "intervention modeling", "mental simulation", "individual-level counterfactual invariance", and "human-level cognition", are used too loosely. Several transitions overpromise: Section 3 promises concrete, domain-specific examples but gives broad categories; Section 6 raises useful AI safety points, but the claim that counterfactual systems are "in principle explainable", needs qualification. The conclusion repeats the thesis instead of resolving the main evidentiary gaps.

Author Response

RESPONSE TO REVIEW 3

Review 3 was the most comprehensive and critical, raising concerns about novelty, scope, evidence quality, and neuroscience interpretation. This review prompted the most substantial revisions. We address each major issue systematically.

Issue 3.1: Novelty Not Sufficiently Established

Reviewer 3 noted: "The manuscript relies heavily on Pearl, Zečević et al., and Jin et al. without clearly establishing what is novel. What the reviewer wants: A sharper original synthesis that goes beyond existing claims, a new taxonomy, a clearer research agenda."

Revision:
We have added an explicit statement of original contribution at the end of the Introduction: "The original contribution of this Viewpoint is to propose counterfactual invariance as a cross-level criterion linking formal causal inference, mental-model-based cognitive simulation, neural scenario construction, and concrete NeuroAI design pathways." [NEW, Introduction, final paragraph]

This frames the specific value: not the claim that LLMs lack counterfactual reasoning (acknowledged as not new), but the systematic connection between formal theory, cognitive mechanisms, neural implementation, and operationalizable AI design criteria. Section 5.1 ("Two Concrete NeuroAI Research Pathways") provides the concrete instantiation of this synthesis.

 

Issue 3.2: Scope Overly Broad - Target System Definition

Reviewer 3 noted: "The manuscript conflates different types of AI systems (LLMs, Current AI systems, Statistical AI, All current machine learning systems) as if interchangeable. Some RL systems approximate Level 2. Some causal discovery systems detect causal structure. The reviewer wants: Define the target class of systems precisely, specify behavioral criteria for Level 2 vs Level 3 competence, address alternative explanations."

Revision:
We have narrowed the scope throughout by introducing the term "text-only LLM reasoners" and using it consistently. [Introduction, Definition section; Abstract]

Changes include:
1. Abstract: Changed "Large language models (LLMs)" to "Text-only large language models (LLMs) used as general-purpose reasoners"
2. Introduction: Added explicit scope statement: "This article does not address artificial intelligence as a whole. Its primary target is large language models used as general-purpose reasoners, especially text-only systems operating without explicit causal models or direct interaction with the world."
3. Throughout: References to "all current machine learning systems" have been removed or qualified. Claims now specify "text-only LLM reasoners" where appropriate.

We have also added explicit discussion of alternative explanations in Section 5.1.2 (MEA pathway), acknowledging that Level 2 and Level 3 failures may reflect benchmark design, poor prompting, insufficient grounding, or architectural limitations—not merely the absence of causal reasoning per se.

 

Issue 3.3: Section 3 Underdeveloped - Expansion with Fourth Category

Reviewer 3 noted: "Section 3 has multiple problems: (a) announces FOUR categories but shows only THREE; (b) examples are asserted rather than demonstrated; (c) claims may be plausible commentary rather than supported argument; (d) no task examples, model outputs, or controls."

Revision:
We have substantially expanded Section 3. The revision includes:

(a) The missing fourth category: "Counterfactual Invariance Under Causally Equivalent Reformulations" [NEW, Section 3.4]. This category provides the strongest diagnostic criterion: stable reasoning should be preserved across paraphrases when the underlying causal structure is constant, and should shift only when causal structure itself changes.

(b) Explicit qualification: We now state that "These failure modes are not individually diagnostic; their relevance is cumulative: together, they motivate the need for controlled tests of counterfactual invariance, such as the diagnostic proposed in §3.4" [NEW, Section 3 opening].

(c) Each failure mode now includes specific literature citations (Jin et al. 2024, Joshi et al. 2024, Zečević et al. 2023, Zou et al. 2023) with brief statements of empirical findings.

(d) Level distinctions: Section 3.2 now explicitly notes that collider reasoning is "primarily a Level 2 failure," and clarifies that "it should therefore not be conflated with Level 3 individual counterfactual inference." [Section 3.2, NEW paragraph].

While we cannot include full model outputs in this Viewpoint format, we have made the claimed failures more concrete by specifying what systems fail at (individual-level counterfactual invariance) and what criterion would demonstrate competence (stability across reformulations).

 

Issue 3.4: Neuroscience Evidence - Over-Interpretation

Reviewer 3 noted: "The paper conflates several related but distinct processes: mental simulation, episodic construction, regret/fictive-error processing, causal learning, counterfactual inference. These need to be separated, not treated as synonymous. Meltzoff (1995) shows intention inference, not formal counterfactual models. DMN and hippocampus findings support episodic simulation but NOT formal counterfactual causal reasoning."

Revision:
We have substantially revised Section 4 to address this concern. Major changes include:

(a) Explicit distinction of component processes: Section 4 now opens with: "Counterfactual reasoning, in its fully explicit human form, is not reducible to a single neural mechanism. It appears to depend on several component processes that are older than language: intention inference, action–outcome learning, episodic construction, fictive evaluation, and model updating." [Section 4, opening, NEW].

This establishes that counterfactual reasoning is composite, and that individual pieces do not constitute the whole.

(b) Meltzoff reinterpreted: We now state: "This shows that intention inference is a component process relevant to later counterfactual reasoning, although it should not be treated as direct evidence for formal counterfactual inference." [Section 4, Meltzoff paragraph, NEW addition].

This addresses the reviewer's concern directly: Meltzoff is now properly framed as one component, not as direct evidence of formal Level 3 reasoning.

(c) DMN and Hippocampus clarified: Section 4.1 now explicitly states: "While episodic simulation and counterfactual reasoning are not identical, they likely share core generative mechanisms." [Section 4.1, opening]. This avoids claiming equivalence while acknowledging potential mechanistic overlap.

(d) New title for Section 4.1: Changed from "Neural Implementation: The Counterfactual Brain" to "Neural Component Processes Supporting Counterfactual Thought" [Section 4.1, NEW heading]. This title explicitly signals that these are components, not a unitary system.

(e) Conclusion modified: Section 7 now acknowledges: "Several evidentiary gaps remain open. It is not yet established which neural circuits specifically support individual-level counterfactual inference, as distinct from episodic construction or fictive error processing." [NEW, Section 7]. This explicitly concedes that the link between component processes and formal Level 3 reasoning is not yet established—a key point the reviewer raised.

 

Issue 3.5: NeuroAI Argument - Historical Oversimplification

Reviewer 3 noted: "The manuscript claims CNN, attention, and RL 'emerged from neuroscience' but this is historically incomplete. CNNs also drew from signal processing. Transformer attention was NOT derived from prefrontal-parietal mechanisms. RL has deep roots in control theory and decision theory; dopamine links came later."

Revision:
We have substantially rewritten Section 5 to provide a more historically accurate and nuanced account. The key change: [NEW, Section 5, paragraph 2-3]. The text now reads:

"CNNs were influenced by Hubel and Wiesel's (1959) discovery of hierarchical receptive fields, but also by advances in signal processing and gradient-based optimization. Reinforcement learning has deep roots in control theory and dynamic programming; the link to dopaminergic prediction-error signals (Schultz et al. 1997) came later as a productive biological convergence, not as a derivation. Transformer attention was not derived from prefrontal-parietal circuits, though structural parallels may still suggest useful hypotheses. In each case, biological findings helped identify useful computational constraints or motifs, without serving as a complete derivation of the resulting AI method."

This acknowledges the reviewer's valid point: neuroscience provided useful constraints and motifs, but was not the sole source of these architectures. The lesson is refined: the productive role of neuroscience is to identify computational problems and functional principles, not to dictate implementations.

 

Issue 3.6: Reference Integrity - Major Corrections

Reviewer 3 noted: "Missing citations, date mismatches, unintended self-citations, incomplete reference notation."

Revision:
Comprehensive reference cleanup performed. [Reference list completely revised] Added missing citations: Allman 2000 (evolutionary neuroscience), Kaas 2013 (brain evolution), Friston 2010 (free-energy principle). All date mismatches corrected and verified against primary sources. The previously included self-citation has been removed because it was not essential to the argument. DOI links have been added where available, with consistent Harvard formatting and proper punctuation.

 

Issue 3.7: Terminological Precision

Reviewer 3 noted: "Terminology used too loosely: counterfactual reasoning, causal reasoning, causal cognition, causal world model, intervention modeling, mental simulation, human-level cognition."

Revision:
We have improved terminological precision throughout:
1. "Counterfactual reasoning" is now more carefully distinguished from "Level 3 reasoning" (the formal characterization).
2. "Causal world model" is now specified as either "formal" (in the sense of Pearl) or "approximated" (in the sense of mental models) where ambiguity could arise.
3. "Human-level cognition" replaced with more specific terms like "robust Level 3 reasoning" or "individual-level counterfactual inference."
4. A new explicit definition statement appears early in the Introduction: "I refer to these models below as 'text-only LLM reasoners.'"

Section 4.2 provides explicit contrast between Pearl's formal definition and Johnson-Laird's cognitive approximation, clarifying the relationship between normative and descriptive approaches.

 

Issue 3.8: Minor Proofreading Issues

Reviewer 3 flagged multiple typos and formatting issues.

Revision: All identified issues corrected:
• "nost likely" → "most likely" [Conclusion]
• "reasons casually" → "reasons causally" [Section 5, now removed via text revision]
• "formal laws theory" → "formal theory" [removed via revision]
• "calculus(Gopnik" → properly formatted [References corrected]
• "neuroscience.:" → "neuroscience." [cleaned]
• Incomplete reference notation "s.d." → proper format [References revised]
• All section headings now follow consistent formatting
• All inline citations now properly formatted with year and author

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

All suggestions and comments have been revised, and this paper is now ready for publication.