Human Authors, AI Verification, Open Challenge

Generative AI has entered academic writing and reviewing at scale, while several quality-control mechanisms of traditional publishing have relied, in part, on a fact that used to hold broadly: producing a respectable text takes effort. Recent evidence from top machine-learning conferences shows that manuscripts containing hallucinated references can pass several reviewers undetected. This is the editorial of Quantum Economics Notes (QEN), a research series. We argue that, in the age of AI, the value of research publishing lies in five things that remain relatively scarce: original evidence, verified claims, transparent sources, accountable judgement and a durable research record. Detection and prohibition can only answer “who wrote it”, and replacing papers with code repositories cannot substitute for verifying claims. QEN therefore reallocates the roles in publishing: authors work with AI on research and writing and are responsible for all content; the QENAI Agent verifies each submission claim by claim and issues a verification report, while the editor decides on acceptance; after publication, readers may challenge an article publicly by submitting a comment. This is an institutional experiment that has not been shown to outperform traditional peer review, and QEN does not describe its articles as peer reviewed. The appendix sets out the full submission guidelines; authors can give this editorial to their own AI to check a manuscript before submission.

1. 1

   Recent evidence from top machine-learning conferences shows that AI is already widely used in peer review, and that manuscripts containing hallucinated references can pass several reviewers undetected.

2. 2

   The design thesis of this editorial is that, in the age of AI, the bottleneck in academic publishing is shifting from producing text to verifying claims. Governance centred on identifying authorship can answer only “who wrote it”, not “does it hold”.

3. 3

   Executable code is valuable evidence but cannot replace the verification of claims: code that runs does not guarantee reproducible results, and reproducible results do not guarantee valid inferences.

4. 4

   QEN reallocates the roles in publishing: authors and AI collaborate on research and writing; review no longer depends on external referees, as the QENAI Agent verifies each submission claim by claim and issues a verification report, while the editor decides on acceptance; after publication, readers may challenge an article publicly by submitting a comment. QEN encourages the use of AI, but authors are responsible for all content.

5. 5

   QEN’s model is a deliberate institutional experiment, not a scheme proven superior to traditional peer review: it assigns systematic verification to the QENAI Agent and judgement to the editor, removes the wait for external referees, and sets a service objective of no more than four weeks of editorial processing time. Public comments keep the research record open to scrutiny after publication. The model cannot replace the scholarly community’s judgement of novelty and importance, and QEN does not describe its articles as peer reviewed.

This editorial proposes the following analytical framework. Several mechanisms of traditional academic publishing have relied, at least in part, on a fact that used to hold broadly: writing a well-structured, coherent and properly referenced text takes considerable time and understanding, so the existence of such a text could serve as a rough signal that the author had actually done the research. Initial screening in peer review and evaluation centred on publication counts have both relied on this signal to varying degrees.

Large language models have weakened this signal, at the writing and the reviewing end at the same time. On the writing side, Sakana AI’s AI Scientist-v2 generated three entirely AI-written manuscripts and submitted them to a workshop at ICLR 2025, where one of them passed review; this was workshop-level review, and the paper was withdrawn before publication, as agreed in advance with the organizers[1]. On the reviewing side, Liang et al. estimate that 6.5–16.9% of review text at four conferences, including ICLR 2024 and NeurIPS 2023, was substantially modified by LLMs[2]; at ICLR 2026, a detection model from the AI-text detection company Pangram Labs classified about 21% of roughly 75,800 reviews as fully AI-generated and found signs of AI involvement in more than half[3]. Both sets of figures are estimates produced by detection models or statistical methods rather than item-by-item human judgements, and they are measured differently, but they point in the same direction.

More importantly, the errors were not caught. GPTZero examined 300 ICLR 2026 submissions: its tool first flagged 90 as containing references that could not be verified online, and manual checks then confirmed that 50 of them contained at least one hallucinated reference. All of these manuscripts had already been assessed by three to five reviewers, and some had average scores as high as 8[4]. Among the 4,841 papers accepted at NeurIPS 2025, GPTZero confirmed a further 100 hallucinated references across 51 to 53 papers (the report’s main text and appendix table differ)[5,6]. Both investigations were carried out by vendors of detection tools, and neither sample is a random draw across disciplines; they do not show that peer review as a whole is unreliable, but they do show that serious errors can pass several reviewers. All of this evidence comes from machine-learning conferences, whose scale and mode of review differ greatly from those of economics journals, so whether it generalizes to other disciplines remains to be tested. There is related evidence from economics: Buchanan et al. found that more than 30% of the economics references produced by GPT-3.5 did not exist, with GPT-4 only slightly better[7]; that study, however, concerns writing rather than reviewing.

In a public lecture at the 2026 International Congress of Mathematicians, Terence Tao anticipated that, if his working hypothesis holds, mathematics will move from “proof scarcity” to “proof abundance”: journals, priority conventions, hiring and promotion criteria, prizes and other institutions were designed under the assumption of scarcity and may behave poorly under abundance[8]. This editorial takes the view that the same holds for academic publishing more broadly: we are moving from text scarcity to text abundance.

QEN is a research series founded on 18 January 2026; publication in QEN is formal publication, not a working paper. This editorial serves two purposes. The main text addresses three questions: under text abundance, what remains relatively scarce in research publishing (Section 2); why two common responses fall short (Section 3); and what QEN does, what it solves and what it does not yet solve (Sections 4–7). The editorial explains the design rationale and does not report how QEN has performed. Appendix A contains QEN’s submission guidelines, written as technical documentation. Authors can give the full text of this editorial to their own AI and ask it to check a manuscript against it, item by item, before submission.

If text is no longer scarce, the value of research publishing must lie elsewhere. The framework proposed here is that generative AI has also lowered the cost of the following five things, but far less than the cost of text, so they remain relatively scarce:

• —
  Original evidence: new data, new measurements, new experiments or new derivations. AI can already produce new derivations and analyses, but these need verification just as human ones do; new observations about the world still depend on real data collection.
• —
  Verified claims: whether a statement is supported by evidence, and to what extent.
• —
  Transparent sources: whether readers can trace every number and every reference to a primary source at reasonable cost. The cases in Section 1 show that, even after several reviewers, checking whether references are genuine can be missed.
• —
  Accountable judgement: who reached which conclusions, and, when something is wrong, who is responsible and how it is corrected.
• —
  A durable research record: whether an article’s content, revision history and corrections can be cited stably over the long term after publication.

None of these depends on how well the text is written; they depend on whether the chains of evidence and responsibility behind the text are complete and open to inspection. In other words, what research publishing must protect in the age of AI is not the process of “humans writing by hand”, but the outcome that “claims can be tested and responsibility can be assigned”.

3.1Detection and prohibition: answering “who wrote it”, not “does it hold”

The first response is to identify and penalize AI-generated content. This faces two difficulties. The first is reliability: Liang et al. tested seven widely used GPT detectors on TOEFL essays written by non-native English speakers and found that more than half were misclassified as AI-generated, while simple prompting strategies easily evaded detection[9]. Using detection results as grounds for sanctions would systematically harm particular groups of authors.

The second is more fundamental: detection asks whether a passage was written by a machine, whereas academic quality asks whether a claim holds. ICLR 2026’s policy already reflects this distinction: it permits the use of LLMs but requires disclosure, makes authors and reviewers ultimately responsible for their contributions, and treats false or misleading claims, fabricated data and hallucinated references as the core violations[10,11]. The conference’s retrospective also acknowledged that, with more than 75,000 reviews, human judgement alone could not identify problems consistently, and that enforcement relied mainly on automatically flagging suspected AI-generated reviews and searching for references to non-existent works. The same retrospective, however, reports that the automated reference checker had a substantial false-positive rate, so each flagged paper was checked manually by area chairs and then programme chairs, involving at least three people, with a channel for appeal[12]. The case shows two things: machines can efficiently flag suspicious references, but people still have to adjudicate what is flagged; and what is truly enforceable, and truly important, to check is references and claims, not authorship.

3.2Replacing papers with code repositories: no substitute for verifying claims

The second response is to replace the PDF with “code that runs”. This captures the value of executable evidence but conflates three levels: code that runs is not guaranteed to reproduce the reported results, and reproducible results do not guarantee that the inferences drawn from them hold. Trisovic et al. re-executed more than 9,000 R files from over 2,000 replication packages on the Harvard Dataverse (published 2010–2020, mostly in the social sciences): 74% failed on the first run, and 56% still failed after automated code cleaning; the authors explicitly treat re-executability as a necessary but not sufficient condition for reproducibility[13].

Moreover, code, like text, can be generated cheaply; and in mathematics, theoretical economics and much qualitative research, the contribution itself does not take executable form. Tao’s suggested rule of thumb is that if authors cannot give a clear, correct and properly attributed expert-level talk on their results, the results should not be published, even if they have been formally verified[8]. Executable code should serve as evidence supporting claims, not as a replacement for them.

4.1Welcoming AI, with full author responsibility

Major publication-ethics bodies and the mathematical community have reached consensus on two points: AI tools cannot be authors, because they cannot take responsibility for the work; and the use of AI must be disclosed truthfully[14,15]. The Leiden Declaration on Artificial Intelligence and Mathematics, released in June 2026 and endorsed by the International Mathematical Union, makes 23 recommendations, including disclosure of tool use and the principle that credit and responsibility belong to humans[8,16]. Nature subsequently published an editorial fully endorsing the declaration’s process and conclusions; the same editorial also relayed the declaration’s position that mathematical results should continue to be published in peer-reviewed venues and that human judgement should be protected when AI is used[17]. On whether to use peer review, QEN makes a different choice, for the reasons given in Section 4.2.

QEN accepts these two principles and goes a step further: QEN does not discriminate against, and indeed encourages, the use of AI tools. Authors may use AI for searching, analysis, programming, drafting and self-checking, and QEN does not use AI-text detection results as grounds for acceptance or rejection. QEN asks for only two things: that authors disclose their use of AI truthfully in the Author Statement, and that they take responsibility for all of the article’s content, including any AI-generated parts. Tao notes that the scenario to be avoided at all costs is one in which authors use AI tools covertly and conceal this to avoid criticism from their peers[8]. By encouraging use and judging only the results, QEN aims to reduce the incentive to conceal.

4.2Reallocating roles: QEN takes on review

Traditional peer review relies on the volunteer labour of external referees, and finding willing referees is itself becoming harder: Publons’ survey of reviewers worldwide found that the number of invitations editors had to send per completed review rose from 1.9 in 2013 to 2.4 in 2017[18]. In emerging interdisciplinary fields the situation may be much worse: as a journal editor, the author has often had to send between 30 and more than 100 invitations to secure two or three referees (personal experience, not systematically recorded). Meanwhile, the evidence in Section 1 shows that referees themselves increasingly use AI in reviewing[2,3].

It should be stressed that these two points do not imply that “AI can replace referees”. That referees use AI shows only that AI has entered the review process; it does not show that AI on its own can detect methodological errors, still less that it can replace the judgement of domain experts. There is evidence that AI can assist review: Liang et al. found that the overlap between GPT-4’s comments and those of human referees was comparable to the overlap between two human referees[19]; and in a randomized study covering more than 40,000 reviews at ICLR 2025, Thakkar et al. found that 27% of referees who received AI feedback updated their reviews accordingly[20]. These studies, however, evaluate AI assisting human referees, not AI reviewing on its own, and Liang et al. themselves caution against using AI comments in place of expert review. Existing research also gives clear warnings: comparing human and AI reviews at ICLR 2026, Baumann et al. found that AI reviews are more homogeneous both within and across papers, and that merely having an LLM rewrite a paper’s style raises AI review scores; they therefore argue that today’s AI systems should not write reviews, or automate judgements that affect acceptance decisions, without rigorous prior evaluation[21]. Tao likewise holds that AI evaluation tools can serve as useful filters that conserve scarce expert attention, but that passing an automated filter is no substitute for acceptance by the community, and he does not believe that human referees can be removed from the publication process[8].

QEN takes these criticisms seriously but deliberately tries a different institutional design, with three safeguards. First, the QENAI Agent only verifies: it checks references, numbers, and the match between claims and evidence item by item, flags methodological concerns, and issues a verification report with a must-fix list; it gives no scores and makes no recommendation to accept or reject. Manipulation aimed at AI scores, such as rewriting a paper’s style to raise them, therefore does not translate directly into acceptance decisions, while blind spots arising from homogeneous AI judgements are offset by the editor’s judgement and by post-publication comments. Second, the editor decides on acceptance and judges whether the methods hold and whether the work is novel and important. Third, QEN does not describe its articles as peer reviewed: under the definition widely used in publication ethics, peer review means obtaining advice from experts in the field who are not part of the editorial team[22]. QEN’s process does not meet this definition, so the accurate description of a QEN article is “AI-verified and editor-approved”. On this basis, QEN reallocates the roles in publishing (Table 1).

Review is divided into three layers: machine-verifiable integrity—references, metadata, numbers, recomputation from code, and the correspondence between claims and sources—is assigned to the QENAI Agent; the evaluation of methods and inference—identification strategies, model assumptions, alternative explanations—is flagged by the QENAI Agent and judged by the editor; novelty, importance, fit with QEN’s scope and the adequacy of the authors’ revisions are judged entirely by the editor.

4.3Key Claims

Another element of QEN’s design is that each article lists 3–5 Key Claims separately, as the core conclusions for which the article is accountable. Key Claims may be the authors’ original findings or synthesized judgements on existing evidence; each must be supported in the main text, and this is exactly what the QENAI Agent checks, claim by claim.

Figure 1 shows the standard process for QEN manuscripts. Manuscripts written by the editorial office follow the same process as external submissions; detailed requirements are given in Appendix A.

1. 01
   Submission (author). Send a Word file to contact@quantecon.ai, ideally after a self-check by your own AI against Appendix A.
2. 02
   Editorial screening (editor). Manuscripts outside the scope, or lacking significant new evidence or new judgement, are returned directly; QEN does not accept incremental contributions.
3. 03
   AI verification (QENAI Agent). The agent verifies references, numbers and Key Claims item by item against the checklist in Appendix A and issues a verification report; it gives no scores and makes no acceptance recommendation.
4. 04
   Response and revision (author). Authors respond to each comment and revise the manuscript.
5. 05
   Editorial decision (editor). Taking into account the verification report and the authors’ response, the editor decides to accept, to request revision and re-verification, or to reject.
6. 06
   Publication (editorial office). A QEN number is assigned, v1.0 is published, and that version is archived on Zenodo.
7. 07
   Comments and updates (readers and authors). Readers may challenge a published article publicly by submitting a comment; authors may publish a new version, which must cite the previous version and explain what has changed and why. Every version is archived on Zenodo.

QEN’s service objective is an editorial processing time of no more than four weeks, excluding the time authors spend on revisions; this is an objective, not a track record. One hard rule applies throughout: references that cannot be verified must be removed.

Numbering and article types. QEN has only three article types: Research Notes, Comments and Editorials. Research Notes are numbered “QEN + year + two-digit sequence”, for example QEN202601; Comments add C and a sequence number to the number of the article commented on, for example QEN202601C1, QEN202601C2; Editorials use the number 00: this editorial is QEN202600, and subsequent editorials are QEN202600A, QEN202600B and so on.

Page structure. Each article contains, in order: title, authors and ORCID iDs, standard citation, abstract, Key Claims, main text, references, Author Statement, Editor Statement and version history. The Editor Statement gives the publication date and the model and number of verification rounds of the QENAI Agent. The website shows the latest version by default, and every version is archived on Zenodo with its own DOI.

Table 2 lists the pain points of traditional publishing that QEN addresses. The column “QEN’s response” describes mechanism design; its effects will have to be tested as QEN operates.

6.1Comments

Post-publication scrutiny is carried out by readers. The “open challenge” in the title refers to this mechanism: any reader may challenge a published article publicly by submitting a comment, numbered by adding C and a sequence number to the number of the original article, for example QEN202601C1. Compared with responding through a separate article, as in traditional journals, there are three differences: comments are numbered under the original article and linked from its page, so that anyone reading the article also sees the challenges to it; comments follow the same fast process as Research Notes, without waiting for external referees; and the original authors may respond by publishing a new version or by submitting a comment of their own. QEN has no procedure for item-by-item corrections: a substantial challenge should be written up as a comment, not sent as scattered fault-finding.

QEN’s model is not without costs. Some of the following problems remain unsolved; others are created by the model itself.

• —
  Limits of AI verification. The QENAI Agent is designed to verify what machines can check, and even its accuracy at that has not yet been systematically tested; whether the methods hold and whether the work is new and important depend on the editor’s judgement and on the reputation QEN builds over time. The QENAI Agent gives no scores, but its must-fix list may still sway the editor’s judgement, so the editor must confirm each item.
• —
  Correlated errors from similar models. The AI used by authors and the QENAI Agent may be based on similar models and make the same mistakes. The QENAI Agent runs on locally deployed open-weight large language models, whose models and versions are updated over time; QEN records the model and version used for each article in its Editor Statement and relies on public comments as a complement, but this risk cannot be eliminated.
• —
  Closed review. QEN does not publish verification reports or authors’ responses; readers see only the final article, the Editor Statement and the version history. The credibility of review therefore rests mainly on the verification items listed in Appendix A and on public comments.
• —
  Security risks of AI review. Studies show that prompt injections hidden in papers can substantially inflate the assessments of AI reviewers, and that simple defences can be circumvented by targeted attacks[25,26,27]. QEN treats every manuscript as untrusted input, detects hidden text and runs code only in isolated environments, and the QENAI Agent gives no scores or acceptance recommendations, but such attacks cannot be ruled out entirely.
• —
  Academic recognition. QEN does not describe its articles as peer reviewed and cannot substitute for peer-reviewed journals in current systems of promotion and academic evaluation; nor does it currently meet the inclusion criteria of the Directory of Open Access Journals (DOAJ), which require, for example, at least two independent reviewers per article[28]. Publication in QEN is formal publication, and the same content may not be submitted to another journal, so authors must choose between QEN’s speed and transparency and the academic recognition of traditional journals. What QEN offers is a service objective of no more than four weeks of editorial processing time, a DOI for every version, open access, and free version maintenance for as long as QEN exists.
• —
  Speed may encourage volume. Faster publication may tempt authors to publish more, reproducing the “text abundance” this editorial criticizes. QEN does not aim for publication volume: editorial screening returns manuscripts without significant new evidence or judgement, and QEN does not accept incremental contributions.
• —
  Loss of writing as training. When humans and AI write together, humans retain topic selection, the framing of claims and the response to review—the stages where judgement is most concentrated—but the value of writing by hand as training for thinking is genuinely reduced. Tao likewise argues that in areas such as education and the training of young mathematicians, the use of AI tools should be restricted quite tightly[8]. QEN is a channel for publishing research records, not a training channel.

Generative AI has made text extremely cheap, and has also made evidence, verification, sources, accountability and records somewhat easier to produce, but far less so than text. QEN uses AI for verification precisely to narrow this gap. Academic norms have converged on the principles that AI cannot be an author, that its use must be disclosed and that responsibility lies with the authors; what is missing is a routine mechanism that puts this consensus into practice for every article and every claim. QEN is a small-scale institutional experiment: it encourages authors to do research with AI, assigns systematic verification to the QENAI Agent and judgement to the editor, and opens post-publication scrutiny to all readers. Its value lies not in producing more text, but in publishing, faster, research records that can be checked, cited, updated and held to account; whether the model works will be judged by its track record.

1. [1]
   Yamada, Y. et al. The AI Scientist-v2: workshop-level automated scientific discovery via agentic tree search. arXiv:2504.08066 (2025). arXiv ↗
2. [2]
   Liang, W. et al. Monitoring AI-modified content at scale: a case study on the impact of ChatGPT on AI conference peer reviews. Proc. 41st Int. Conf. Mach. Learn. (ICML), PMLR 235, 29575–29620 (2024). Link ↗
3. [3]
   Emi, B. Pangram predicts 21% of ICLR reviews are AI-generated. Pangram Labs Blog, 18 November 2025. Link ↗
4. [4]
   Esau, P. GPTZero finds over 50 new hallucinations in ICLR 2026 submissions. GPTZero, 6 December 2025. Link ↗
5. [5]
   Shmatko, N. GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers. GPTZero, January 2026. Link ↗
6. [6]
   Ansari, S. Compound deception in elite peer review: a failure mode taxonomy of 100 fabricated citations at NeurIPS 2025. arXiv:2602.05930 (2026). arXiv ↗
7. [7]
   Buchanan, J., Hill, S. & Shapoval, O. ChatGPT hallucinates non-existent citations: evidence from economics. The American Economist 69, 80–87 (2024). DOI ↗
8. [8]
   Tao, T. Mathematics in the age of AI. arXiv:2608.16753 (2026). arXiv ↗
9. [9]
   Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. & Zou, J. GPT detectors are biased against non-native English writers. Patterns 4, 100779 (2023). DOI ↗
10. [10]
    ICLR 2026 Program Chairs. Policies on large language model usage at ICLR 2026. ICLR Blog, 26 August 2025. Link ↗
11. [11]
    ICLR 2026 Program Chairs. ICLR 2026 response to LLM-generated papers and reviews. ICLR Blog, 19 November 2025. Link ↗
12. [12]
    ICLR 2026 Program Chairs. A retrospective on the ICLR 2026 review process. ICLR Blog, 31 March 2026. Link ↗
13. [13]
    Trisovic, A., Lau, M. K., Pasquier, T. & Crosas, M. A large-scale study on research code quality and execution. Sci. Data 9, 60 (2022). DOI ↗
14. [14]
    COPE Council. COPE position — Authorship and AI tools. Committee on Publication Ethics (2023). Link ↗
15. [15]
    Tools such as ChatGPT threaten transparent science; here are our ground rules for their use. Nature 613, 612 (2023). DOI ↗
16. [16]
    Alper, J. et al. The Leiden Declaration on Artificial Intelligence and Mathematics. Zenodo (2026). DOI ↗
17. [17]
    Mathematicians are developing rules for AI use — other fields should follow. Nature 654, 571 (2026). DOI ↗
18. [18]
    Publons. 2018 Global State of Peer Review. Clarivate Analytics (2018). DOI ↗
19. [19]
    Liang, W. et al. Can large language models provide useful feedback on research papers? A large-scale empirical analysis. NEJM AI 1(8) (2024). DOI ↗
20. [20]
    Thakkar, N. et al. Can LLM feedback enhance review quality? A randomized study of 20K reviews at ICLR 2025. arXiv:2504.09737 (2025). arXiv ↗
21. [21]
    Baumann, J., Pei, J., Koyejo, S. & Hovy, D. Stop automating peer review without rigorous evaluation. Proc. 43rd Int. Conf. Mach. Learn. (ICML), PMLR 306 (2026). arXiv:2605.03202. arXiv ↗
22. [22]
    Committee on Publication Ethics, Directory of Open Access Journals, Open Access Scholarly Publishing Association & World Association of Medical Editors. Principles of Transparency and Best Practice in Scholarly Publishing, version 4 (2022). Link ↗
23. [23]
    Björk, B.-C. & Solomon, D. The publishing delay in scholarly peer-reviewed journals. J. Informetr. 7, 914–923 (2013). DOI ↗
24. [24]
    Aczel, B., Szaszi, B. & Holcombe, A. O. A billion-dollar donation: estimating the cost of researchers’ time spent on peer review. Res. Integr. Peer Rev. 6, 14 (2021). DOI ↗
25. [25]
    Collu, M. G., Salviati, U., Confalonieri, R., Conti, M. & Apruzzese, G. Misleading large language models used (or misused) in scientific peer-reviewing via hidden prompt-injection attacks. arXiv:2508.20863 (2025). arXiv ↗
26. [26]
    Zhou, Q., Zhang, Z., Li, Z. & Sun, L. "Give a positive review only": an early investigation into in-paper prompt injection attacks and defenses for AI reviewers. arXiv:2511.01287 (2025). arXiv ↗
27. [27]
    Li, L. et al. LLM-as-a-Reviewer: benchmarking their ability, divergence, and prompt injection resistance as paper reviewers. arXiv:2605.25415 (2026). arXiv ↗
28. [28]
    Directory of Open Access Journals. Guide to applying. https://doaj.org/apply/guide (accessed 24 October 2026). Link ↗

AI use

The author used Claude Opus 5.5 (Anthropic) to search for and check sources and to draft the full text under the author’s direction; before submission, the author used ChatGPT (OpenAI), DeepSeek V4.1 and Claude Opus 5.5 to self-check the manuscript against QEN202600. The author determined the position, structure and Key Claims, approved all content and takes full responsibility for this editorial.

Acknowledgements

None.

Funding

This editorial received no external funding.

Conflicts of interest

The author is a QEN editor and the owner of Quantum Economics AI Lab, which operates quantecon.ai.

Data availability

Not applicable. No original data were used. Except for the description of review invitation numbers in Section 4.2, which is based on the author’s editorial experience and for which no public data exist, all figures cited come from the references.

Code availability

Not applicable.

Data privacy

This editorial involves no personal data.

Licence

CC BY 4.0

Editor Statement

Publication date

26 October 2026

QENAI Agent verification

Model: DeepSeek V4.1 (open-weight, deployed locally); protocol: QEN202600 v1.0, Appendix A4; rounds and dates: 3 rounds on 25 October 2026

Current v1.0 — 26 October 2026 — First published.