Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Average
Originality of the paper: Yes, but limited
Adequacy of the bibliography: No

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Limited
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

Review of Temporal Neuro-Symbolic Reasoning: From Architectures to Verifiable and Auditable Systems.
Capsule Summary
This paper analyzes temporal neurosymbolic reasoning through logical formalisms,
neurosymbolic integration, interpretability, and evaluation protocols.
It suggests a structured roadmap organized into maturity levels aimed at guiding the field toward verifiable systems.
The paper is clearly written, uses PRISMA 2020 guidelines, and contributes in organizing integration paradigms.
However, the survey contains a critical fundamental technical and historical flaw: it systematically omits the fundamental foundational body of work that established temporal neurosymbolic reasoning and learning as a research field.
Very specifically, the works developed in a dozen papers by Garcez, Lamb, Borges, and collaborators starting at NIPS 2003 and at least until 2014 is ignored.
This is not a light omission.
It concerns the papers that first demonstrated connectionist temporal logic, temporal knowledge learning in recurrent networks, temporal rule extraction, and agentic deployment of temporal neurosymbolic systems in applicable simulators (see IJCAI 2011, etc).
A survey claiming to cover 1980–2025 cannot be credible or published while ignoring this foundational literature.

Contributions
The integration typology on Logic => Network, Network => Logic, and Network <=> Logic is a useful guideline.
The work on STL robustness semantics and connection to differentiable temporal reasoning is technically sound.
The signed-distance interpretation in Figure 8 and the pipeline of Figure 9 are helpful.
The roadmap in Table 2 linking evaluation criteria, benchmark requirements, and control capabilities over maturity levels is clear.
These contributions are real but do not compensate for the omissions described above and further described below.

Critical Needed Change for a Survey Paper from 1980:
The omission of the foundational work directly related to the core of the paper;
these previous works were published in leading conferences and journals (NIPS/IJCAI/AAAI/IEEE TNN, etc, so they are easily accessible).
The survey fails to acknowledge, cite, or situate the pioneering work of Garcez, Lamb, Borges, and collaborators that created the field of temporal neurosymbolic AI.
Their work directly and clearly addressed the survey central research objectives.
Their work and was the first to do so in a grounded, implemented, and empirically validated manner (published not only at AI and Machine Learning venues, but also at the flagship International Conference on Software Engineering, IEEE TNN, and other venues).

Garcez and Lamb (NeurIPS-formerly NIPS 2003): established the foundational correspondence between temporal logic operators.
They already back in 2003 included past and future LTL modalities. They showed that key agent knowledge representation problems (muddy children puzzle, and others) could be reasoned about and learned in a neurosymbolic architecture. This is the foundational work on neurosymbolic temporal logic learning and reasoning. The work builds a connectionist architecture with soundness guarantees. This work was extended and the journal version with full results and extended examples has been published in Neural Computation: Artur S. d'Avila Garcez, Luís C. Lamb:
A Connectionist Computational Model for Epistemic and Temporal Reasoning. Neural Computation 18(7): 1711-1738 (2006).
The survey discusses connectionist temporal logic integration as if these foundational results did not exist. This is a crucial omission.

Lamb, Borges, and Garcez (AAAI 2007): introduced Sequential Connectionist Temporal Logic (SCTL), proving a correct correspondence between a fragment of LTL extended with past operators and NARX recurrent networks. The paper demonstrated temporal synchronization and learning (using the classical Dining Philosophers Problem as a testbed) within a unified connectionist model. It proved translation soundness and extracted temporal rules from trained networks.
This directly instantiates the Network <=> Logic integration mode the survey presents as a contemporary development (they did so back in 2007), with formal guarantees than most systems surveyed.
Not citing this work in a 2026 survey on neurosymbolic temporal reasoning is a scientific and academic omission or/and methodological error.

Borges, Garcez, and Lamb (IJCNN 2007): extended SCTL specifically to past temporal operators, demonstrating their connectionist implementation. The survey discusses past operators as a formal tool without acknowledging this prior implementation.

Borges, Garcez, and Lamb (ICANN 2010): provided the full neuralsymbolic cycle for temporal knowledge. It shows the translation of temporal logic programs into NARX networks. It shows how to perform (temporal) learning from examples and properties, and extraction of revised temporal logic programs. The survey Table 4 Network => Logic begins in 2016 with DT-STL, misrepresenting the history of the field by several years.

Borges, Garcez, Lamb, and Nuseibeh (ICSE 2011): applied SCTL (neurosymbolic) to software engineering. It integrates it with the NuSMV model checker in a cycle of verification, counter-example generation, and neural learning. It achieved in 2011 what the survey roadmap presents as a medium-term goal.
It is the first paper that showed that counterexamples cycle to neurosymbolic learning engine is a powerful methodology. Several papers are now using exactly the same principle for model refinement.
See e.g. David Harel, Assaf Marron, Ariel Rosenfeld, Moshe Y. Vardi, Gera Weiss:
Labor Division with Movable Walls: Composing Executable Specifications with Machine Learning and Search (Blue Sky Idea). AAAI 2019: 9770-9774.

Borges, Garcez, and Lamb (IEEE Transactions on Neural Networks 2011): is the journal consolidation of the SCTL neurosymbolic temporal learning and reasoning methodology and program. It presents formal semantics, soundness, gradient-based temporal learning, pedagogical extraction, and model checking integration, that can be applied and has been applied in software engineering and beyond. Its omission from a survey of 1980–2025 coverage cannot be justified by any selection criterion.

de Penning, Garcez, Lamb, and Meyer (IJCAI 2011): built directly on SCTL to implement the first deployed agentic system capable of online temporal reasoning and learning. Directly anticipates the survey long-term autonomy goals stated in the roadmap.

de Penning, Garcez, Lamb, Stuiver, and Meyer (IJCNN 2014): deploys the temporal neurosymbolic cognitive agent in intelligent transportation, reducing CO2 emissions. Demonstrates real-world scalability over a decade before the survey roadmap identifies as a future goal.

Survey omissions and claims:
These omissions materially impact the validity of specific claims.
The survey situates STLnet (2020) and T-LEAF (2021) as early exemplars of tight neurosymbolic integration.
However, while SCTL achieved provably sound integration of temporal semantics with differentiable neural computation in 2007. It was extended at least until 2014.
Therefore, chronological starting points of all appendix tables are incorrect given the omitted literature.
The roadmap presents isolated and a posteriori inspectable temporal reasoning as a short-term goal to be built, however the SCTL framework implemented exactly this in 2010. The medium-term goal of architectures directly integrating temporal constraints with abductive correction was achieved by Borges et al. 2011 through counter-example-driven iterative learning with the model checker NuSMV.
Presenting these as open directions, while ignoring their prior achievement, misrepresents past scientific contributions.

The survey cites Garcez and Lamb (2020) as a conceptual reference, while ignoring the same authors previous temporal neurosymbolic work.
The 2020 paper summarizes and builds upon the omitted foundations.
The PRISMA selection process may have a systematic bias toward post 2015 publications, even though the authors stated 1980–2025 timeframe. Maybe the bias/use of certain techniques to write the survey inadvertently penalized foundational papers with a terminology different from the authors use of terms.

SUGGESTED DECISION:
Major Revision Required
In a survey, the omission of the Garcez-Lamb-Borges foundational program (published in leading venues, therefore accessible) is a structural deficiency.
It directly affetcs the historical framing, the chronological history/inventory, the roadmap of open problems, and the survey's credibility as a comprehensive paper.

REQUIRED CHANGES: The literature mentioned is based on foundational papers, not random selections.
The original contributions of such an important field must be acknowledged in a comprehensive peer-reviewed academic survey.

Garcez and Lamb, NeurIPS 2003: should be credited as establishing the first soundness-guaranteed connectionist temporal logic framework. This is one of the foundational papers of the field.
Lamb, Borges, and Garcez, AAAI 2007: should be credited as establishing foundations and evaluation of SCTL and the bidirectional integration of LTL and NARX networks with formal proofs and validation in temporal synchronisation, reasoning and learning.
Borges, Garcez, and Lamb, IJCNN 2007: connectionist treatment of past temporal logic operators learning.
Borges, Garcez, and Lamb, ICANN 2010: complete neural-symbolic cycle for temporal knowledge including rule extraction.
Borges, Garcez, Lamb, and Nuseibeh, ICSE 2011: first integration of temporal neuro-ymbolic learning with a formal model checker in a software engineering pipeline, using the now deployed technique of fully automated model refinement using counter-examples to inform learning.
Borges, Garcez, and Lamb, IEEE Transactions on Neural Networks 2011: definitive formal treatment of SCTL with soundness proofs and empirical validation using software engineerig research benchmarks.
de Penning, Garcez, Lamb, and Meyer, IJCAI 2011: first sound deployed agentic temporal neurosymbolic system.
de Penning, Garcez, Lamb, Stuiver, and Meyer, IJCNN 2014: first real-world application of temporal neurosymbolic agents to reduce CO2 emissions.

On non-classical logics: Garcez and Lamb wrote a piper titled "Neural-symbolic systems and the case for non-classical reasoning', where they argued the case for several non-classical logics in neurosymbolic and machine learning. They explicitly mentioned temporal, intuitionistic, and modal logics as key to the development of sound neurosymbolic systems.
Artur S. d'Avila Garcez, Luis C. Lamb:
Neural-Symbolic Systems and the Case for Non-Classical Reasoning. We Will Show Them! (1) 2005: 469-488.
Their subsequent monograph, called a landmark in neurosymbolic AI by Gary Marcus consolidated the case for non-classical reasoning and learning:
Artur S. d'Avila Garcez, Luis C. Lamb, Dov M. Gabbay:
Neural-Symbolic Cognitive Reasoning. Cognitive Technologies, Springer 2009. Easily accessible and highly cited literature.
This work presents ''self-contained presentation of neural network models for a number of logics, including modal, temporal, and epistemic logic." Thus, temporal and non-classical neurosymbolic methods are already mentioned in the book.

Table corrections:
Table 3 (Logic => Network) include SCTL-based approaches from 2007 onward as early examples of the integration mode.
Revise Table 4 (Network => Logic) to begin its chronological information to 2010, correcting the current 2016 starting point which misrepresents the field's history by six years.
Revise Table 5 (Bidirectional) to include SCTL from 2007 as the first bidirectional framework with formal soundness guarantees.
Include a row or comment in each table explicitly positioning the SCTL neurosymbolic system relative to current approaches on key dimensions: formal guarantees, differentiability, scalability, and empirical validation (SCTL has been validated in practice and their system pioneered a cycle used today in software engineering research).

Roadmap revisions
Revise short-term foundations section to acknowledge that inspectable temporal reasoning with rule extraction was shown first in 2010–2011.
Reframe the short-term goal as reproducibility and standardization of this capability rather than its initial establishment.
Revise the medium-term integration section to acknowledge that iterative verification-learning cycles with model checkers were implemented back in 2011, and reframe the goal.
Revise the long-term autonomy section to acknowledge that agentic online temporal neurosymbolic learning was demonstrated in 2011, and deployed in 2014, distinguishing genuinely open challenges from already-achieved milestones and proofs of concepts.

Historical academic/scientific framing:
Rewrite the introduction to fairly and historically represent that bidirectional integration of temporal logic and recurrent neural networks with soundness proofs was achieved in 2007-2011, and not in the 2020s.
Revise the integration typology section to position the SCTL program as the first formally sound instantiation of the Network <==> Logic.
Revise claims about STLnet (2020) and T-LEAF (2021) as pioneering integration approaches to reflect their relationship to prior foundational work from Neurips/NIPS 2003, AAAI2007, IEEE TNN 2011, Neural Computation, etc by Garcez et al.
Methodology adopted:
Review if PRISMA somewhat penalized foundational papers with pre-2015/2020 terminology. Describe any corrective measures.
Verify the 1980–2025 period coverage against the actual distribution of retained papers; correct bias towards recent publications.
The specific required changes (add the listed papers to tables/roadmap/introdction, corrected chronology and open challenges, potential PRISMA bias are necessary for the paper to be academic credible/publishable as a 2026 survey.

Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

The paper advances three primary claims: (i) that difficulties in temporal neuro-symbolic reasoning are primarily structural rather than purely computational; (ii) that dominant evaluation practices are inadequate for assessing trajectory-level coherence; and (iii) that a maturity roadmap can guide the field toward auditable systems. The second claim is particularly well supported, while the first, although insightful, seems somewhat overstated and not supported by a quantified discussion. The third constitutes a novel synthesis that connects architectural, evaluative, and governance perspectives.

The papers gives a taxonomy of integration paradigms (Logic → Network, Network → Logic, Network ↔ Logic)
In terms of originality, the taxonomy of integration paradigms (Logic → Network, Network → Logic, Network ↔ Logic) is not entirely new, similar distinctions appear in Garcez & Lamb (2020), which the authors cite. However, the paper extends this framework productively by decomposing the Logic → Network category into four mechanistic sub-schemes (regularization, latent-space constraints, action filtering, and automaton integration). This refinement has clear practical value for guiding architectural choices. Additionally, the Pareto-style framing of performance versus coherence trade-offs provides a useful conceptual lens, although it is presented as an illustrative visualization rather than a formal result.

The survey of temporal logics in Section 1 (LTL, STL, Allen’s algebra, hybrid logics) largely covers well-established material. However, I think that he survey leaves several important areas underexplored. Probabilistic temporal logics receive only little attention treatment, despite their relevance for reasoning under uncertainty. Similarly, neuro-symbolic approaches to temporal planning and action languages (e.g., STRIPS or PDDL with learned components) are absent, even though they represent a significant strand of research at the intersection of reasoning and decision-making. Finally, recent work on temporal commonsense reasoning in LLms is only lightly addressed. Moreover, the survey would benefit from more systematic citation of foundational sources for the different temporal logics, (eg., they cite mainyl the book by Clarke et. al).

The roadmap is one main paper’s contributions. Its three-horizon structure (auditability → integration → autonomy) seems logically coherent, and Table 2 gives a good overview by linking metrics, benchmarks, architectures, and capabilities. However, it seems developed. The short-term horizon lacks specificity (e.g., no concrete definition or examples of temporal perturbations). The medium term confuses architectural milestones with evaluation milestones, which should be treated separately. The long-term horizon is largely aspirational and not clearly falsifiable. Finally, although governance is highlighted in the abstract, it is not meaningfully integrated into the roadmap.

Also, the treatment of LLMs is underdeveloped. Given the rapid growth of work on temporal reasoning in LLMs since 2022, the paper’s engagement with this area is relatively limited. While benchmarks such as TISER and LTLBench are cited, the paper does not analyze why LLMs struggle with temporal reasoning, nor does it clearly articulate what neuro-symbolic approaches offer beyond prompting or chain-of-thought strategies.

Reviewer has chosen to be Anonymous

Overall Impression: Excellent

Content:
Technical Quality of the paper: Excellent
Originality of the paper: Yes
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Excellent

Detailed Comments:

Overall Impression

This survey delivers on its ambitious promise: it successfully positions temporal neuro-symbolic (NeSy) reasoning as a principled pathway toward verifiable and auditable AI systems. A particular strength is that the authors avoid the common trap of producing yet another catalog of techniques. Instead, they construct a coherent intellectual framework that exposes structural limitations of current approaches—not merely algorithmic shortcomings—and propose a forward-looking roadmap with concrete maturity levels. This is, in my view, precisely the kind of survey the field needs: one that synthesizes rather than merely enumerates, and that connects technical architectures to governance requirements.

---

Technical Quality: Excellent

The paper demonstrates exceptional technical depth across three dimensions:

1. Formal rigor: The treatment of temporal logics (LTL, STL, interval logics, hybrid systems) goes beyond superficial mentions. Section 1 precisely distinguishes dense-time vs. discrete-time formalisms, explains robustness semantics in STL with concrete mathematical examples (Figure 8), and addresses undecidability results in interval logics (Halpern & Shoham, 1991). This isn't hand-waving—it's graduate-level formal methods made accessible.

2. Architectural taxonomy: The decomposition of integration modes (Logic→Network, Network→Logic, Network↔Logic) in Section 2 provides genuine analytical leverage. Figure 5's four schemes (A–D) expose a continuum of constraint enforcement from soft regularization to hard filtering. This directly serves the paper's verifiability thesis: stronger integration modes enable stronger guarantees.

3. Critical analysis of trade-offs: The paper doesn't shy from uncomfortable truths. It explicitly documents the vanishing-gradient problem in Gödel t-norms (p. 18, citing Klement et al., 2000 and Krieken et al., 2022), the semantic gap between Boolean and fuzzy satisfaction (p. 24), and the scalability cliff as temporal horizons extend (p. 24). These aren't footnotes—they're central arguments justifying the roadmap's staged maturity levels.

One caveat: The PRISMA methodology (Figure 1) filters 2,576 publications down to 73 using a composite score threshold (1.75). While transparent, the paper doesn't justify why this threshold yields a representative sample. A sensitivity analysis (e.g., "what changes if we lower the threshold to 1.5?") would strengthen methodological credibility.

---

Originality: High

This isn't incremental survey work. Three contributions stand out:

1. Structural critique: The paper reframes limitations as inherent to time itself—"non-locality of time, irreversibility of constraints, dependence on complete history" (p. 2)—rather than engineering gaps to be solved with bigger GPUs. This philosophical grounding elevates the survey beyond technical reporting.

2. Roadmap with maturity levels: Table 2 isn't boilerplate "future work." It links evaluation criteria (short-term: auditability), benchmarks (synthetic perturbations), architectures (separable chains), and governance capabilities (distinguishing perception vs. reasoning errors) into a conditional progression. Each level is a prerequisite for the next—this is systems thinking rarely seen in surveys.

3. Evaluation reframing: The authors identify that dominant practices "favor point-wise performance and do not assess global consistency of trajectories" (abstract). Figures 13–14 operationalize this critique using aeIOU/TAC metrics and Pareto-front analysis—turning an abstract concern into an actionable methodology.

Limitation: The roadmap's long-term horizon ("self-regulated systems capable of revising rules") remains speculative. The paper acknowledges this, but could better anchor it in existing work on lifelong learning (e.g., Thrun's foundational texts from the 1990s, which are currently absent).

---

Adequacy of Bibliography

The PRISMA methodology transparently filters 2,576 → 73 papers with explicit scoring (threshold 1.75). The survey covers core temporal logics (LTL, STL, interval logics) and key NeSy architectures (PyReason, TILP, NeSyA), and includes 2024–2025 preprints showing currency (ONSEP, TISER, LTLZinc).

However, three bibliographic gaps weaken the manuscript's foundation and should be addressed:

1. Justification for a temporal-specific survey. While the paper cites foundational NeSy works (Garcez & Lamb, 2020), it does not explicitly articulate why temporal reasoning warrants a dedicated survey rather than being subsumed under general NeSy taxonomies. This justification is crucial because many NeSy challenges are exacerbated by time's sequential nature. The authors could strengthen the motivation by explicitly contrasting temporal NeSy with static NeSy along three axes: (i) non-locality – temporal constraints can span arbitrarily long horizons, unlike static rules which are typically local; (ii) compositionality – events combine via ordering operators (e.g., *since*, *until*) that have no static analog; and (iii) evaluation – trajectory-level coherence matters more than point-wise accuracy. Adding 2–3 paragraphs clarifying these distinctions would justify the survey's scope and help readers understand why temporal reasoning is not merely a "subsection" of NeSy but a distinct subfield with unique structural challenges.

2. Underrepresentation of constraint-based temporal reasoning. The survey's treatment leans heavily on temporal logics (LTL, STL) and Allen's interval algebra. However, it omits several foundational frameworks from the constraint-based reasoning community that are directly relevant to NeSy integration:
- Point Algebra (Vilain & Kautz, 1986): This provides a simpler, tractable fragment of Allen's relations and is widely used in scheduling. Its absence is notable because many NeSy applications (e.g., robotics planning) operate with point-based timelines.
- Event Calculus (Kowalski & Sergot, 1986): This is the de facto standard for narrative reasoning and action effects in symbolic AI. Its absence limits the paper's coverage of "narrative" understanding, a growing application area for LLMs (as the paper itself notes with TISER and ONSEP).
- Temporal Constraint Networks (TCNs): While Dechter et al. (1990) is cited, the survey does not discuss modern TCN literature (e.g., Simple Temporal Networks, STNUs) that deals with controllability and robustness under uncertainty—concepts that map directly to STL robustness semantics (p. 18).

Incorporating these frameworks would not require expanding the survey dramatically; a single paragraph in Section 1.3 ("Temporal representation models") contrasting interval-based, point-based, and event-based representations would suffice. This would also strengthen the roadmap's "medium-term" goal of handling delays and causal dependencies (Table 2), for which TCNs provide well-understood verification mechanisms.

3. Lifelong learning foundations. The roadmap's long-term vision of "self-regulated systems capable of revising rules" would benefit from anchoring in established lifelong learning literature (e.g., Thrun, 1998), which provides theoretical and algorithmic foundations for knowledge retention and revision under non-stationarity.

---

Abstract and Introduction

Abstract: The abstract clearly states the problem ("structural limitations"), the method ("unified structuring of integration paradigms"), and the contribution ("roadmap organized into maturity levels") without hype. Crucially, it foregrounds the verifiability/auditability mission—not buried in paragraph 4. One improvement: explicitly name the three integration modes (Logic→Network, etc.) to signal the taxonomy's centrality.

Introduction: The introduction correctly frames temporal reasoning as trajectory validation—not instantaneous prediction (p. 2). It crisply contrasts neural approaches ("statistical correlations") vs. symbolic ("explicit constraints") without caricature. The motivation for NeSy fusion emerges naturally from this tension.

Weakness: The roadmap promise ("structured into successive maturity levels") appears only in the final paragraph. The introduction should preview Table 2's three horizons earlier—readers need to know why they're traversing 40 pages of formalisms. A single sentence after the problem statement ("We address this through a three-stage roadmap...") would provide essential scaffolding.

---

Organization: Satisfactory with Structural Friction

The flow—Foundations → Integration → Interpretability → Evaluation → Roadmap—is logical. The PRISMA diagram (Figure 1) upfront establishes methodological credibility. Appendix tables (3–5) provide valuable cross-referencing without cluttering the narrative.

Structural friction: Section 3 ("Interpretability") and Section 4 ("Evaluation") suffer from topic bleed. The aeIOU/TAC metrics (Figure 13) belong in Evaluation, but the abstraction chain (Figure 12) is fundamentally about interpretability. A sharper boundary—e.g., "Section 3: Making reasoning inspectable; Section 4: Measuring reasoning validity"—would reduce cognitive load. Additionally, the roadmap (Section 5) repeats points from Sections 2–4 without sufficient synthesis.

---

Language and Presentation

English level: Satisfactory. No grammatical errors impede comprehension.

Minor stylistic issues:
- Occasional French-influenced phrasing: e.g., "reasoning trajectories" appears throughout; "inference traces" (the term used by PyReason, which they cite) would be more natural in English AI literature.
- Overuse of "constitute"—varying with "form," "represent," or active verbs would improve readability.
- Long sentences in Section 4 (e.g., p. 29, a 58-word sentence beginning "Unlike static tasks...") could be split for clarity.

These are stylistic preferences—not barriers to understanding. No native-speaker revision is required.

---

Overall Presentation: Excellent

Figures 5, 8, 11, and 14 are pedagogical masterpieces: they transform abstract concepts (integration modes, robustness semantics, explanation formats, Pareto trade-offs) into visual intuition without oversimplifying. Table 1's comparative taxonomy enables quick orientation.

Most importantly: every visual serves the roadmap thesis. Figure 5's integration schemes directly map to Table 2's maturity levels; Figure 14's Pareto front justifies why short-term auditability must precede long-term autonomy. This coherence between exposition and argument elevates the presentation beyond decorative illustration.

---

Recommendation

Accept with minor revisions.

Summary of required revisions:
1. Bibliography: Add 2–3 paragraphs contrasting temporal vs. general NeSy challenges to justify the survey's scope.
2. Bibliography: Expand coverage of constraint-based temporal reasoning (Point Algebra, Event Calculus, modern TCN literature) and anchor the long-term roadmap in lifelong learning foundations (e.g., Thrun).
3. Introduction: Preview the three-stage roadmap (Table 2) earlier to provide readers with structural scaffolding.
4. Organization: Sharpen the boundary between Section 3 (interpretability) and Section 4 (evaluation) to reduce topic bleed.
5. Methodology: Provide a brief justification for the PRISMA threshold choice (1.75) to enhance replicability.