From Deliberately Constructed Systemic Representations of the World of Activity to Dialogue: A Five-Level Communication Architecture for Strong Language-Based ASI

arXiv:2610.32869v1 [cs.AI] 23 Oct 2026

From Deliberately Constructed Systemic Representations of the World of Activity to Dialogue: A Five-Level Communication Architecture for Strong Language-Based ASI Maxim Osovsky

Adam University

Bishkek

A B S T R A C T. In working communication with models and agents of strong language-based

ASI, users must be able to understand the basic representations of the world, activity, the economy, and economic organization on which those models and agents rely. The central thesis of this paper is that, at present, a language model’s context is fundamentally incapable of presenting the foundations of utterance formation for examination and verification: these foundations are established in advance by acquired schemas, a systemic representation of the world of activity, and the connections accepted as true within it. Neither context nor trained LLM weights, in themselves, attest to the presence of deliberately constructed systemic representations of the world of activity. The paper proposes five levels of content organization for strong language-based ASI: deliberately constructed systemic representations of the world of activity, a situation as a fragment of those representations, its narrative description, the full sense of a message, and text. A situation’s schema-model is not identical to a story about it. A recipient can reconstruct the full sense of a message from an abbreviated text through jointly constructed and agreed-upon representations and the situation of joint activity. Reverse work is integral to the architecture: proceeding from the full sense of the message through an extended description of the situation and its schema-model to the reconstruction of systemic foundations. The theoretical foundations draw on G. P. Shchedrovitskiy’s systems methodology, A. Korzybski’s structural differential, and I. A. Mel’čuk’s Meaning–Text theory. The author’s contribution is a specific systemic-activity organization of content that makes it possible to examine and adjust the foundations of a message and to control the consequences of changes to them. Further research concerns the construction of large graphical models of activity on reconstructible systemic foundations, as distinct from the generation of text from context. Keywords: Ai, AGI, ASI, SI, Artificial Intelligence, Artificial Superintelligence; Artificial General Intelligence, Super Intelligence, activity-based communication; Meaning–Text Theory; generative grammar; discourse comprehension; systems approach; systemic representations.

1

A RT I C L E C O N T E N T S 1.

P o s i t i o n : W h a t I s t h e Fo u n d a t i o n o f a L a n guage Model’s Response?

1.1. Token Prediction and the Representation of Sense . . . . . . 1.2. Context, Internal Memory, and Graph Representations . . .

3 14 15

2. Five Levels of Organization

20

3. Architectural Implications

27 29

3.1. Understanding a Message and the Criterion of ASI Adequacy 3.2. Testing the Architectural Proposal . . . . . . . . . . . . . .

33

Conclusion

35

D E C L A R AT I O N S

36

2

1 . P o s i t i o n : W h a t I s t h e Fo u n d a t i o n o f a L a n g u a g e Model’s Response?

In human communication, a response is grounded in acquired schemas, a systemic representation, and the connections accepted as true within it, rather than in the mere presence of some context. P. G. Shchedrovitskiy formulates the distinction between presenting a representation and mastering it: “The world must not only be presented in the forms of ideal representations and ontological pictures; it must also be brought within the sphere of human practical activity in its real forms”.1 In this paper, a foundation is an organization of representations that establishes distinctions between processes and structures, specifies their connections, and determines the permissible content of a message in joint activity. What comes first in forming a description of activity: the context from which structural and systemic representations emerge, or already acquired, albeit incomplete, representations that determine how that context is constructed? The proposed formulation requires consideration of both directions: elaborating a representation into a linguistic description and reconstructing or restructuring a representation from a description. These transitions are not identical to compressing and unpacking text. Distinctions between processes and structures, and the connections between them, must be preserved and checked within the representation, and what is lost, introduced, or changed in the transition must be identified. The task of language AI is to master this bidirectional work, not merely to continue text. If the necessary systemic representations have not been acquired, understanding requires work to construct them. Such work may take place in communication, but it cannot be reduced to reading the message and may take a long time. This paper’s contribution concerns the systemic-activity organization of the foundations of a message, a separate narrative description of the situation, and control of transformations between the five levels. The proposed five levels specify the organization of message formation and understanding within thinking and activity in which ASI participates. Its theoretical foundation is considered 1.

P. G. Shchedrovitskiy, Философия развития и конструктивная психология [Philosophy of Development and Constructive Psychology], in Педагогика самоопределения и гуманитарные практики [The Pedagogy of Self-Determination and Humanistic Practices] 38 (Barnaul, 1999) (pp. 38–48; item 9), https://shchedrovitskiy.com/ph ylosophiya-razvitiya-i-konstruktivnaya-psychologiya. Unless otherwise indicated, translations of quotations from Russian sources are by the author.

3

in Section 2. Mel’čuk restricts the task of semantic decomposition to “determining the applicability of LUs to things and situations already identified” and excludes the identification of the thing or situation itself.2 This distinction separates linguistic description from the representations from which a situation is cut out and on whose basis it is organized. Reconstructing a situation from a linguistic description already presupposes schemas through which what is described can be recognized. Discussing the recognition of another participant in activity, P. G. Shchedrovitskiy observes: “And when he encounters the fact that others are not like him, yet cannot recognize the other person (he has no other schema, so to speak), a whole host of interpersonal and content-related conflicts begins”.3 What matters for this paper is the connection between recognition and the availability of a schema. Accordingly, the question of a language model’s reconstruction of a situation first requires establishing which acquired schemas it uses and which connections it accepts as true when interpreting the description. Anna Elashkina explicitly discusses the distinction between ontological foundations and their textual description in relation to AI: “Whereas, previously, such an ontology effectively determined the architecture of a software system in logical systems, a working ontology now has the status of a text which, although it may have some priority, remains merely a text in a neural network and certainly not its architecture”.4 For this paper, the distinction between functions is essential: a description of an ontology can be conveyed through context, but 2.

I. Mel’čuk, Explanatory Combinatorial Dictionary, in Open Problems in Linguistics and Lexicography 225, § 2.1.1 (G. Sica ed., Polimetrica, Monza, 2006) (pp. 225–355), https://olst.ling.umontreal.ca/pdf/Mel%27c%CC%8Cuk_2006_ECD.pdf.

3.

P. G. Shchedrovitskiy, Введение в синтаксис и семантику графического языка СМД-подхода [Introduction to the Syntax and Semantics of the Graphical Language of the Systemic Thought-Activity Approach], Electronic Library of the G. P. Shchedrovitskiy Institute for Development Foundation (Oct. 8, 2010) (seventh semester; lecture 38 in the overall cycle; Moscow, Academy of National Economy; transcript), https://www.fondgp.ru/old/lib/mmk/58/10-10-08_Schedrovitckij_P.G._shematiz atciya_-_XXXVIII.doc.

4.

Anna Elashkina, Искусственный интеллект просит подумать [Artificial Intelligence Asks Us to Think], Открытые системы. СУБД [Open Systems. DBMS], no. 4 (2024), https://doi.org/10.51793/OS.2024.81.42.002; https: //www.osp.ru/os/2024/04/13059028.

4

the presence of that description is not identical to the organization of a system in which accepted distinctions and connections accepted as true serve as foundations for the formation of a message. The subject of this paper is the conversational machines of AI systems that form linguistic messages in joint activity. A person or a human–machine system participates in constructing this world and in communication within it. Acquired schemas take shape in a particular linguistic environment and develop historically. The acquisition of a representation cannot be reduced to reading or processing descriptions. A picture of the world is acquired primarily through actions and activity. N. Ya. Marr connected manual speech with manual thinking and emphasized its significance “as a means of organizing production, behavior, and worldview”.5 A person acts, works with material, interacts with other people, and masters means of activity. Internalization concerns all of this, not only text. Speech, terminology, meanings, and sense are important, but they do not exhaust the acquisition of representations. The paper draws on the author’s experience of organizing collective thinking through the “Schematization” technology and on observations of game participants’ speech, schemas, activity, and communication.6 Between 2011 and 2026, the author conducted more than 350 games involving up to 300 people at a time, using the method of “organizational activity games” described by Marina Liborakina.7 Participants jointly describe and schematize the complex systems of activity in which they are involved, considering their existing organization in relation to their historical formation and possible future. In this work, the author observes thinking through the interrelations of speech, schemas, activity, 5.

N. Ya. Marr, Язык и мышление [Language and Thinking], in 3 Избранные работы [Selected Works] 90, 106 (1934) (pp. 90–121; lecture delivered on June 26, 1931; quotation at 106; electronic publication by CRECLECO, University of Lausanne), https://crecleco.seriot.ch/textes/MARR-31b/txt.html.

6.

M. Osovsky, Технология игры «Схематизация» [The Technology of the “Schematization” Game], Medium (June 15, 2018), https://osovsky.medium.com/game-5438c 730a15e.

7.

M. Liborakina, A bridge between past and future, in Games and Simulations to Enhance Quality Learning 41 (Simulation and Gaming Yearbook, vol. 4, Psychology Press, 1996) (pp. 41–48; ISBN 9780749418663), https://books.google.com/books?i d=4meAFnexIMoC&dq=Moscow+Methodological+Circle&pg=PA48.

5

and communication. Differences in the identification of elements, connections, and processes become objects of discussion and of coordination among participants’ representations. This practice provided the grounds for formulating a task for language AI: the formation of a message must rest on representations of the activity being described and maintain their connection to the content being expressed. The study of thinking in organized joint work has a foundation in the tradition of organizational activity games. P. G. Shchedrovitskiy considers such a game a method of comprehensive socio-psychological research and emphasizes: “A ‘game’ of this kind is, in essence, real life activity, but compressed in time and space and taking place under the artificial conditions of game organization”.8 The organization of a game makes it possible to isolate the processes of thinking and communication occurring within it for investigation. For this paper, this is important as a basis for observing how an utterance takes shape within the participants’ activity. Yrjö Engeström develops a related research formulation in the approach of formative interventions and the practice of the Change Laboratory. Here, the unit of analysis is a collective activity system. In his study of work in a Finnish hospital, Engeström shows how participants redefine the initial problematic situation and the conceptual instrument being used as they transform their practice.9 This approach provides grounds for considering the development of representations together with changes in the activity in which they are used. In Cognition in the Wild, Edwin Hutchins investigates the thinking of a ship’s navigation team as the work of a system that includes participants and the means they use.10 He traces the propagation and transformation of representations 8.

P. G. Shchedrovitskiy, Организационно-деятельностная игра как метод комплексного социально-психологического исследования [The Organizational Activity Game as a Method of Comprehensive Socio-Psychological Research], in Программно-целевой подход и деловые игры [The Program-Target Approach and Business Games] 36 (Novosibirsk, 1982) (pp. 36–39; electronic publication), https://shchedrovitskiy.com/igra-kak-metod-issledovaniya/.

9.

Y. Engeström, From Design Experiments to Formative Interventions, 21 Theory & Psychology 598 (2011) (no. 5; pp. 598–628), https://doi.org/10.1177/095935431141 9252.

10. E. Hutchins, Cognition in the Wild (MIT Press, Cambridge, MA, 1995) (author’s introduction), https://pages.ucsd.edu/~ehutchins/citw.html.

6

across different media and treats communication as part of the organization of collective task performance. What matters for the present formulation is the shift of the unit of analysis from an isolated individual to joint activity, within which representations are formed and used. Kirsikka Vaajakallio and Tuuli Mattelmäki consider collaborative design games as a means of organizing participants’ cooperation.11 Their work situates the use of games in the research context of jointly constructing design content. Together with the tradition of organizational activity games, formative interventions, and research on distributed cognition, this approach provides a basis for examining speech and schemas within collective thinking. The author’s own observations and these research approaches provide the grounds for developing the proposed five-level organization of content. Comparison with Chomsky clarifies the subject of the proposed architecture. Chomsky and his coauthors define language as “a computational operation occurring in the mind of an individual, independent of its possible communicative use.”12 This paper examines the systemic-activity foundations of message formation: historically acquired schemas for cutting out a situation, representations of the world of activity, processes of constructing and transforming content in communication, and the activity system in which the means of thinking are formed and used. This distinction concerns human utterances as well as AI. Word meanings change over time and differ across communities. The persistence of a word therefore does not attest to the persistence of the representations on whose basis it is used. The author’s position is that every language results from the importation of technologies—situations of activity—together with the terms that explain them. From this perspective, the transfer of words must be investigated together with the mastery of that activity and the transformation of the receiving community’s representations. Chomsky, by contrast, considers language a result of the inter11. K. Vaajakallio & T. Mattelmäki, Design Games in Codesign: As a Tool, a Mindset and a Structure, 10 CoDesign 63 (2014) (no. 1; pp. 63–77; Aalto University researchportal record), https://doi.org/10.1080/15710882.2014.881886; https://research.aal to.fi/en/publications/design-games-in-codesign-as-a-tool-a-mindset-and-a-structure /. 12. J. J. Bolhuis et al., Language: UG or Not to Be, That Is the Question, 13 PLOS Biology e1002063 (2015) (no. 2; coauthors: I. Tattersall, N. Chomsky, R. C. Berwick), https: //doi.org/10.1371/journal.pbio.1002063.

7

action between the initial state of the language faculty and experience.13 What is decisive for this paper is the conceptual representation (schema) of a specific situation of activity cut out from a systemic representation of reality as a fragment. For the argument of this paper, it is important that transferring a term does not substitute for mastering the situation of activity that it explains. Knowing a word does not mean acquiring a schema: one must master the distinctions between processes, structures, and their connections with which that word is used in the receiving community. Another problem is that scientific and professional communities assign their own specialized meanings to the same words. An accepted definition is incorporated into the norms of professional usage and determines which assertions and actions a community considers permissible within its subject domain. If AI’s task is to express a given sense, the representation of that sense is primary, not the probability of the next word. A probabilistic generator may select a formulation, but the permissibility of its content must be determined separately from its linguistic probability. There are academic grounds for this distinction. In Meaning–Text theory, a semantic graph represents content that permits different linguistic realizations.14 Iordanskaja and Mel’čuk separately consider the conceptual representation of a situation from which the semantic representation of a sentence is constructed.15 These transitions have earlier foundations: Kulagina and Mel’čuk distinguish the levels of situation and sense and discuss the possibility of intermediate levels 13. N. Chomsky, New Horizons in the Study of Language and Mind at 4 (Cambridge University Press, 2000) (ch. 1, p. 4; official excerpt), https://doi.org/10.1017/CBO9 780511811937; https://assets.cambridge.org/97805216/51479/sample/978052165 1479wsc00.pdf. 14. I. A. Mel’čuk, Towards a Linguistic “Meaning ⇔ Text” Model, in Trends in Soviet Theoretical Linguistics 33 (1973) (pp. 33–57), https://doi.org/10.1007/978-94-010 -2536-2_2. 15. L. Iordanskaja & I. Mel’čuk, Minimodel of Semantic Synthesis of Russian Sentences, 44 Lingvisticæ Investigationes 101 (2021) (no. 1; pp. 101–136), https://doi.org/10.107 5/li.00058.ior.

8

between them.16 In a later account, Mel’čuk places the Meaning–Text model within a global model of linguistic behavior that includes extralinguistic reality and conceptual and semantic representations.17 His 1981 work also considers the semantic representation of an utterance realized by a sequence of sentences.18 Cousin explicitly sets the task of controlling the expression of a message through formally implemented structures of this theory.19 In research that does not refer directly to Mel’čuk, Narayan and colleagues introduce a content plan before generation,20 while the Large Concept Models team proposes processing at a higher semantic level followed by decoding into text.21 The separation of the substantive organization of a message from its linguistic realization has engineering precedents in classical natural language generation. Reiter and Dale distinguish document planning, microplanning, and surface realization. Planning determines the content and structure of a document, microplanning selects the linguistic means of expressing it, and surface realization transforms the resulting specification into text. The communicative goal, the model of the recipient, and discourse history are taken into account in forming a 16. O. S. Kulagina & I. A. Mel’čuk, Automatic Translation: Some Theoretical Aspects and the Design of a Translation System, in Machine Translation 137, 145–146 (A. D. Booth ed., North-Holland, Amsterdam, 1967) (pp. 137–171; situation and meaning levels, intermediate representations at 145–146), https://olst.ling.umontreal.ca/pdf /Mel%27c%CC%8Cuk_1967_Kulagina.pdf. 17. I. Mel’čuk, Représentation sémantique [Semantic Representation], in La théorie SensTexte. Concepts-clés et applications [Meaning–Text Theory: Key Concepts and Applications] 21, 47 (S. Marengo ed., L’Harmattan, Paris, 2021) (pp. 21–47; diagram of the global model of linguistic behavior at 47), https://olst.ling.umontreal.ca/pdf /Mel%27c%CC%8Cuk_2021_RS%C3%A9m.pdf. 18. I. A. Mel’čuk, Meaning-Text Models: A Recent Trend in Soviet Linguistics, 10 Annual Review of Anthropology 27, 34 (1981) (pp. 27–62), https://olst.ling.umontreal.ca/p df/Mel%27c%CC%8Cuk_MTM_RecentTrends_1981.pdf. 19. M. Cousin, Dependency Structures Representation: Meaning Text Theory’s DeepSyntax Encoding With Abstract Categorial Grammars, MoL 163 (2025) (pp. 163– 175), https://aclanthology.org/2025.mol-1.13/. 20. S. Narayan et al., Conditional Generation with a Question-Answering Blueprint, 11 TACL 974 (2023) (pp. 974–996), https://aclanthology.org/2023.tacl-1.55/. 21. LCM team, L. Barrault et al., Large Concept Models: Language Modeling in a Sentence Representation Space, arXiv:2412.08821 (2024), https://arxiv.org/abs/2412.088 21.

9

message.22 In the proposed architecture, this separation is supplemented by the systemic foundations for cutting out a situation and by the distinction between its description and the full sense of a particular message. On the side of text comprehension, the model developed by Kintsch and van Dijk is an important precursor. The authors distinguish the semantic representation of a text—the textbase—from a model of the situation to which the text refers. The situation model includes events, actions, and participants, as well as relevant knowledge and experience that are not necessarily expressed in the message. They also identify the need for a model of the communicative context that represents speech acts and their underlying intentions.23 This distinction separates the reconstruction of what was said from the reconstruction of the situation described and the purpose of the message. This paper also sets the task of reconstructing the systemic representations from which that situation was cut out. In reconstructing G. P. Shchedrovitskiy’s position, P. G. Shchedrovitskiy defines sense as follows: In this position, “sense” is defined as the configuration of connections and relations between different elements of the situation of activity and the speech process. This configuration is created and reconstructed by a person understanding the text of a message.24 22. E. Reiter & R. Dale, Building Natural Language Generation Systems chs. 3– 6 (Cambridge University Press, 2000) (chs. 3–6; official chapter pages), https: //www.cambridge.org/core/books/abs/building-natural-language-generatio n-systems/document-planning/6401F096F58AC31DE37A8668BF07BAEF; https://www.cambridge.org/core/books/abs/building-natural-language-gener ation-systems/microplanning/FEB90C3A816AD6A873A1BE85B69A9D0B; https://www.cambridge.org/core/books/abs/building-natural-language-generation-s ystems/surface-realisation/D45A2AC42657177D950874E6322D2609. 23. T. A. van Dijk & W. Kintsch, Strategies of Discourse Comprehension at 12, 338 (Academic Press, New York, 1983) (pp. 12, 338; author’s copy), https://discourses.org/w p-content/uploads/2022/06/Teun-A-van-Dijk-Walter-Kintsch-1983-Strategies-O f-Discourse-Comprehension.pdf. 24. P. G. Shchedrovitskiy, Идея «мысли=коммуникации»: смысл vs содержание [The Idea of “Thought=Communication”: Sense vs. Content] §13, Pyotr Shchedrovitskiy’s Website (Jan. 27, 2025) (last visited Oct. 19, 2026), https://shchedrovitskiy.com/idey a-myisli-kommunikatsii-smyisl-vs-soderzhanie/.

10

Research on AI also addresses the opacity of the sources of an answer. Neeman and colleagues show that the entanglement of parametric and contextual knowledge makes it difficult to establish the source of an answer and propose generating answers from these sources separately.25 The political orientation of LLM responses and the absence of political views made explicit in advance represent another aspect of the opacity of the foundations of an utterance. Westwood, Grimmer, and Hall show that users perceive the responses of most of the LLMs studied as left-leaning, and that an instruction to remain neutral reduces perceived slant.26 Rozado likewise identifies the political orientation of conversational models’ responses and demonstrates that it can be changed through fine-tuning on politically oriented data.27 Possible causes include the effects of pretraining, fine-tuning, evaluators’ preferences, and system instructions. An orientation in the utterance is observable, but which representations of society, the state, the economy, and activity underlie it has not yet been established. A change in perceived slant achieved through an instruction must therefore be distinguished from the reconstruction and revision of the foundations of message formation. The requirement to disclose the foundations of decisions also matters for AI auditing. Glenn identifies the need for the “Ability for humans to check why the AGI is doing something questionable or undesirable.”28 Applied to a linguistic response, this raises questions not only about sources of information but also about the value-based foundations for selecting and organizing content: which representations of society and activity are accepted, what is considered permis25. E. Neeman et al., DisentQA: Disentangling Parametric and Contextual Knowledge with Counterfactual Question Answering, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 10056 (2023) (pp. 10056–10070; coauthors: R. Aharoni, O. Honovich, L. Choshen, I. Szpektor, O. Abend), https://doi.org/10.18653/v1/2023.acl-long.559. 26. S. J. Westwood et al., Measuring Perceived Slant in Large Language Models Through User Evaluations (2025) (coauthors: J. Grimmer, A. B. Hall), https://modelslant.com /paper.pdf. 27. D. Rozado, The Political Preferences of LLMs, 19 PLOS ONE e0306621 (2024) (no. 7), https://doi.org/10.1371/journal.pone.0306621. 28. J. Glenn, Artificial General Intelligence: Issues and Opportunities at 9 (Foresight on Demand: Foresight towards the 2nd Strategic Plan for Horizon Europe, February 2023) (§ 2.5, p. 9), https://millennium-project.org/wp-content/uploads/2023/05/E C-AGI-paper.pdf.

11

sible, desirable, or impermissible within them, and how these distinctions determine the message. The proposed architecture sets the task of making such foundations available for examination, without reducing their identification to an assessment of the orientation of the finished text. In the OntoAgent architecture, the ontology, fact repository, and linguistic lexicon constitute explicitly organized resources for reasoning, modeling, and communication.29 McShane and Leon separate the agent’s formation of message content from its linguistic generation.30 Contemporary developments of this approach envisage using an LLM as an instrument within a system that retains control over representations and operations on them. Nirenburg and colleagues describe the acquisition of new lexical and ontological knowledge with the aid of an LLM whose proposals are checked against formal ontological constraints.31 Oruganti and colleagues compare OntoAgent with six LLMs within the HARMONIC robotics system, including cases in which the models receive equivalent procedural knowledge. In the task investigated, the availability of such knowledge did not eliminate errors in checking the system’s own state of knowledge, diagnosis, and action selection.32 These results distinguish the provision of information to a language model from the organization of a system in which information is used to form decisions and messages. In Wikontic, an “ontology” is a technical schema of entity types and per29. M. McShane & S. Nirenburg, A Knowledge Representation Language for Natural Language Processing, Simulation and Reasoning, 6 International Journal of Semantic Computing 3 (2012) (no. 1; pp. 3–23), https://doi.org/10.1142/S1793351X1240001 6. 30. M. McShane & I. Leon, Language Generation for Broad-Coverage, Explainable Cognitive Systems, arXiv:2201.10422 (2022), https://arxiv.org/abs/2201.10422. 31. S. Nirenburg et al., LLMs as Knowledge Workers, 8 Proceedings of the AAAI Symposium Series 512 (2026) (no. 1; pp. 512–520; coauthors: M. McShane, S. Oruganti, J. English, C. Arndt), https://doi.org/10.1609/aaaiss.v8i1.42583. 32. S. Oruganti et al., Why Cognitive Robotics Matters: Lessons from OntoAgent and LLM Deployment in HARMONIC for Safety-Critical Robot Teaming, arXiv:2603.26730 (2026) (coauthors: S. Nirenburg, M. McShane, J. English, M. Roberts, C. Arndt, R. Parasuraman, L. Sentis), https://arxiv.org/abs/2603.26730.

12

missible connections.33 In our view, however, it is neither an ontology in the philosophical sense nor a systemic representation of activity that distinguishes processes, structures, and their connections in the sense of this paper. Nevertheless, working with corpora of texts and their fragments produces impressive results in language models, although the grounds of this effectiveness still require explanation. The achievements are particularly conspicuous in relatively restricted formal languages, especially programming languages, in which models create and reproduce code. The successful solution of a local task may rely on a limited working ontology of the situation and need not require reconstruction of an entire system of representations of the world. Understanding the foundations of the solution requires reconstructing the representations the model used and the connections between them. The next task for research and conversational LLMs is to reconstruct the ontological foundations of linguistic descriptions and to construct large graphical models of economic organization, the economy, construction, and softwaresystem development. Reproducing a text about the architecture of software or an industrial enterprise does not attest to the reconstruction of that system’s schema-model. Such a schema must preserve the distinctions between processes and structures and the connections between them and serve as a means of working with them, not merely as an image. L. A. Byzov demonstrates the problem of moving from a sequential description to an integrated representation using historical material: “Speech therefore inevitably destroys the actual relations between events occurring simultaneously and at different times (or between changing states).” He then observes that “special mental efforts and auxiliary means (chronological comparison tables) are needed to reconstruct from chaos and fragments any reasonably integrated pic33. A. Chepurova et al., Wikontic: Constructing Wikidata-Aligned, Ontology-Aware Knowledge Graphs with Large Language Models, arXiv:2512.00590 (2025) (version 2 of Jan. 29, 2026; coauthors: A. Bulatov, M. Burtsev, Y. Kuratov), https://arxiv.org/ht ml/2512.00590v2.

13

ture of a historical period”.34 The order of exposition does not coincide with the organization of the situation described. Understanding requires reconstructing the connections between events and states within a single representation. For AI, the crucial question is: which acquired representations and means does a language model use for such reconstruction? Processing a sequence of text and producing a coherent response do not, in themselves, demonstrate that this work has been performed. In this paper, ASI (superintelligence) means a model capable of reconstructing the systemic organization of activity: schemas of systems and subsystems, processes and structures, synchronized timelines of activity and their connections, and the ontological pictures from which it acts. Reconstruction means constructing representations with which the model can work, rather than merely generating their linguistic descriptions or graphical images. Universality in task performance, the quality of text, and the ability to modify its own code are not, in themselves, sufficient indicators of ASI in this sense. 1 . 1 . To k e n P r e d i c t i o n a n d t h e R e p r e s e n t a t i o n o f S e n s e

In the GPT approach, next-token prediction is the technical task of pretraining. In their GPT-2 paper, the authors hypothesized that a language model with sufficient capacity would begin to learn tasks represented in natural-language sequences in order to predict those sequences better.35 This hypothesis justifies using language modeling as a means of acquiring a broad range of capabilities. The GPT-4 technical report explicitly states that the model is based on the Transformer and pretrained to predict the next token in a document, after which it is 34. L. A. Byzov, Докладная записка об общих проблемах графического языка и необходимости организации Института графического языка [Memorandum on the General Problems of Graphical Language and the Need to Establish an Institute of Graphical Language], note 71, proposition 6, Electronic Edition of «Мы нашли новый язык» [“We Have Found a New Language”] (note 71: abstracts of the paper «История как предмет графического изображения» [“History as an Object of Graphical Representation”], proposition 6), https://docs.google.com/document/d/1 70WgJMGu_j_tL4NGumjNcmMZvroC4sTBtNH3h4X-HIQ/edit?tab=t.0. 35. A. Radford et al., Language Models are Unsupervised Multitask Learners, OpenAI (2019) (coauthors: J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever), https://cdn. openai.com/better-language-models/language_models_are_unsupervised_multitask _learners.pdf.

14

fine-tuned using human feedback.36 The training task explains the selection of a probable sequence, but the probabilistic selection of a formulation does not establish the permissibility of its content. If a language model is given the sense of a message, it is necessary to check separately whether the selected formulation expresses precisely that sense and preserves its connections and conditions. The transition from token prediction to the linguistic realization of a message therefore requires checking the specified content, not merely assessing the probability of the text. Using a probabilistic token generator for the linguistic realization of a message is, in itself, an extremely controversial experiment with unpredictable consequences. Its use does not eliminate the need to organize the foundations of content separately and to check that the given sense is preserved in the text. 1 . 2 . C o n t e x t , I n t e r n a l M e m o r y, a n d G r a p h Representations

Developers also use operational definitions of context. OpenAI defines context as follows: “Here, context refers to the total window of tokens (input + output) that the model can attend to at once.”37 Anthropic defines it as follows: “Context refers to the set of tokens included when sampling from a large-language model (LLM).”38 When discussing external notes, the authors specify: “These notes get pulled back into the context window at later times.”39 An agent system selects instructions, message history, materials, and tool results for its next call to the model.40 Distributed vector representations emerge during context processing. To accelerate generation, a KV cache retains the attention keys and values computed for positions already processed, allowing them 36. OpenAI, GPT-4 Technical Report, arXiv:2303.08774 (2023), https://arxiv.org/abs/23 03.08774. 37. E. Okcular, Context Engineering — Short-Term Memory Management with Sessions, OpenAI Cookbook (Sept. 9, 2025), https://developers.openai.com/cookbook/exa mples/agents_sdk/session_memory. 38. Anthropic, Effective Context Engineering for AI Agents (Sept. 29, 2025), https://ww w.anthropic.com/engineering/effective-context-engineering-for-ai-agents. 39. Anthropic, Effective Context Engineering, supra note 38, § Structured note-taking. 40. Anthropic, Effective Context Engineering, supra note 38.

15

to be reused without recomputation.41 The cache is a numerical state of computation, not an explicitly labeled inventory of participants, predicates, conditions, and unknowns. The Claude Code documentation describes the agent’s work as a recurring cycle of three phases: gathering context, taking action, and verifying results. The results of tool use return to the model and influence its selection of subsequent actions.42 Such a cycle organizes the model’s interaction with its environment. The five levels proposed in this paper address a different task: distinguishing systemic foundations, the conceptual representation of a situation, its narrative description, the full sense of a message, and text. Gathering context, acting, and checking results do not, in themselves, determine the systemic representations from which a situation has been cut out or how its content has been transformed into a message. The agent cycle and the five-level organization of content describe different aspects of AI’s work: the former describes the sequence of interaction with the environment, whereas the latter describes the organization of representations and transitions between them. In the xAI materials examined, the emphasis is on knowledge retrieval and context length. The announcement of Grok states: “We are working on methods that can discover and retrieve information whenever it is needed.”43 In the announcement of Grok-1.5, increased “memory capacity” is explicitly linked to expanded context length: “an increased memory capacity of up to 16 times the previous context length.”44 An agent’s long-term memory may store histories, summaries, plans, and results in external files or databases. Compression of materials and structured notes help it continue long-running tasks.45 In Memorizing Transformers, the authors propose a different memory mechanism: “we extend language models with the ability to memorize the internal 41. Hugging Face, How caching works, Transformers Documentation (last visited Oct. 18, 2026), https://huggingface.co/docs/transformers/en/cache_explanation. 42. Anthropic, How Claude Code works, Claude Code Documentation, https://code.cla ude.com/docs/en/how-claude-code-works. 43. xAI, Announcing Grok (Nov. 3, 2023), https://x.ai/blog/grok. 44. xAI, Announcing Grok-1.5 (Mar. 28, 2024), https://x.ai/news/grok-1.5. 45. Anthropic, Effective Context Engineering, supra note 38.

16

representations of past inputs.”46 Additional retrievable memory differs from knowledge in the weights and from the current context window. When selecting, compressing, and reusing materials, it is necessary to preserve the level of representation to which they belong, their connections and constraints, and the grounds of the information accepted. Research by Google Research and DeepMind distinguishes context from parametric knowledge. In their study of T5 and PaLM, Li and colleagues formulate a hierarchy of preferred sources: “relevant context > model’s parametric knowledge > irrelevant context” (equations (1)–(2)).47 In Racing Thoughts, the authors observe: “Out-of-order contextualization can result in incorrect or inconsistent LLM generations.”48 OpenAI research considers acquired internal representations separately from the current input material. Gao and colleagues write: “Since language models learn many concepts, autoencoders need to be very large to recover all relevant features.”49 The presence of trained LLM weights does not, in itself, attest to the presence of deliberately constructed systemic representations of the world of activity. This requires establishing whether the model distinguishes processes, structures, and their connections and uses these distinctions to cut out a situation and form and understand a message. The opacity of content selection also appears in the generation of fictional stories. In Elias in the Lighthouse, Again?, 88.3% of 20,000 stories generated by four language models contained at least one of 11 persistently recurring words denoting names, occupations, and settings. Yet stories containing such words accounted for only about 3.8% of stories in the OLMo fine-tuning data exam46. Y. Wu et al., Memorizing Transformers, abstract, ICLR (2022) (coauthors: M. N. Rabe, D. Hutchins, C. Szegedy), https://arxiv.org/abs/2203.08913. 47. D. Li et al., Large Language Models with Controllable Working Memory, Findings of the Association for Computational Linguistics: ACL 2023, 1774, 1775 (2023) (pp. 1774–1793; coauthors: A. S. Rawat, M. Zaheer, X. Wang, M. Lukasik, A. Veit, F. Yu, S. Kumar), https://aclanthology.org/2023.findings-acl.112/. 8 4 . M. A. Lepori et al., Racing Thoughts: Explaining Contextualization Errors in Large Language Models, NAACL 3020 §1 (2025) (pp. 3020–3036; coauthors: M. C. Mozer, A. Ghandeharioun), https://aclanthology.org/2025.naacl-long.155/. 49. L. Gao et al., Scaling and Evaluating Sparse Autoencoders, abstract, OpenAI, arXiv:2406.04093 (2024), https://arxiv.org/abs/2406.04093.

17

ined. The frequency of these elements in the corpus does not explain their predominance in responses as a simple reproduction of the distribution of training examples.50 A user’s request for a story does not specify particular characters and situations, yet models systematically introduce similar content. The existence of a trained mechanism and an available corpus therefore does not yet reveal the foundations on which a language model cuts out a situation and selects the content of a message. A computational graph describes dependencies between operations, attention connections describe interactions between positions during processing, and a semantic graph describes the sense-related connections of a linguistic description. Attribution graphs reconstruct the influence of internal features through an approximate replacement model.51 In an Anthropic study, Templeton and colleagues observe: “Many features are multilingual (responding to the same concept across languages) and multimodal (responding to the same concept in both text and images).”52 In their study of Claude 3.5 Haiku, the authors write: “We find the model uses a mixture of language-specific and abstract, language-independent circuits.”53 A 2026 study by researchers from Google DeepMind, Google Research, and universities distinguishes the encoding of novel semantics from their use in openweight models: “open-weights LMs struggle to deploy representations of novel semantics that are defined in-context, even if they encode these semantics in their latent representations.”54 In the topology experiment, the authors report: “LMs 50. S. Hamilton & D. Mimno, Elias in the Lighthouse, Again? Diagnosing Low Diversity in LLM Stories, arXiv:2605.26492 (2026) (version 1), https://arxiv.org/html/2605. 26492v1. 51. E. Ameisen et al., Circuit Tracing: Revealing Computational Graphs in Language Models, Anthropic (Mar. 27, 2025), https://transformer-circuits.pub/2025/attributi on-graphs/methods.html. 52. A. Templeton et al., Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, introductory part, Anthropic (May 21, 2024), https://transformer-c ircuits.pub/2024/scaling-monosemanticity/index.html. 53. Anthropic, On the Biology of a Large Language Model, introduction (2025), https: //transformer-circuits.pub/2025/attribution-graphs/biology.html. 54. M. A. Lepori et al., Language Models Struggle to Use Representations Learned InContext, abstract, arXiv:2602.04212 (2026) (version 2; coauthors: T. Linzen, A. Yuan, K. Filippova), https://arxiv.org/html/2602.04212v2.

18

achieve substantially higher accuracy when given the explicit description than when they need to infer the topology in-context.”55 These studies distinguish the presence of material, its internal representation, and its functional use. The recent ThoughtComm study investigates communication between AI agents through models’ internal representations. Contextualized agent states are transformed into a shared latent representation from which individual continuous prefixes are formed for subsequent generation. A channel for transmitting internal representations thus supplements textual exchange.56 The question of transmitting such representations must be distinguished from that of the organization of the object of inquiry they represent: which processes, structures, and situations they represent, and on what foundations these have been identified. Taken together, these sources distinguish contextual tokens, parametric knowledge, additional memory mechanisms, and internal conceptual representations. Investigating internal mechanisms does not replace establishing the foundation that determined the formation of a particular message. Maintaining the content of a whole text has a separate foundation in P. G. Shchedrovitskiy’s text: Developing what the speaker intends to express and planning the speech, selecting individual sentences and words, changing the order of exposition, and, for more complex texts, creating a system of references and footnotes indicate that, even when an individual sentence is being written down or spoken, both the preceding and the following fragments of the text must be maintained in memory in accordance with its subject-matter content.57 This requirement to maintain the subject-matter content of a whole text provides a transition to distinguishing the levels of its organization and transformation into a message. 55. Lepori et al., Language Models Struggle to Use Representations Learned In-Context, supra note 54, §4. 56. Y. Zheng et al., Thought Communication in Multiagent Collaboration, 38 Advances in Neural Information Processing Systems (2025) (coauthors: Z. Zhao, Z. Li, Y. Xie, M. Gao, L. Zhang, K. Zhang), https://proceedings.neurips.cc/paper_files/paper/2025/ hash/b2b502c3629beadda06311386d2c6f73-Abstract-Conference.html. 57. Shchedrovitskiy, Идея «мысли=коммуникации» [The Idea of “Thought=Communication”], supra note 24, §17.

19

2. Five Levels of Organization

The author proposes five levels of content organization on the basis of experience organizing collective thinking in the “Schematization” game. During the game, participants jointly construct and discuss representations of complex activity systems. An individual participant proceeds from acquired systemic representations of the world of activity, describes them in a graphical language, cuts out a particular situation from those representations as a fragment, elaborates its schema into a graphical description, determines together with other participants the content needed for a message, and expresses it in graphical schemas. Dialogue accompanies this communication at a fairly primitive level. Indeed, very primitive dialogue messages devoid of substantive content can support the participants’ actual communication. When understanding another participant’s message, the work proceeds in the reverse direction: from the graphical schemas presented to reconstruction of the message’s content, the extended description of the situation, its conceptual representation, and the systemic representations from which that situation was cut out. The proposed architecture transfers this organization of message formation and understanding to language-based ASI by distinguishing systemic representations, the conceptual representation of a situation, its narrative description, the full sense of a message, and text. What is transferred comprises the distinctions within content and the directions of work on it, not the assignment of graphical expression to a single level: graphical schemas are available at each of the five levels, not only at the level of situation representation. In the linguistic architecture, an extended graphical description of a situation is transferred into a sequential verbal description at the third level. What is transferred is the function of content organization—the elaboration and comparison of the representations of the processes, structures, and connections being described—not the graphical means itself. Graphical schemas remain means of work at every level. The third level describes the situation, whereas the fourth selects and organizes from that description the content required for a particular message in light of the communicative goal. The first level: deliberately constructed systemic representations of the world of activity. Communication in joint activity requires distinguishing the processes taking place, the structures in which they occur, and their connections. Without these distinctions, it is impossible to establish which organization of activity has been accepted in cutting out a situation and selecting a message. The representations use a graphical language that distinguishes processes, functional 20

and morphological structures, connections, material, price, value, and evaluation, as well as past, present, and future. This systemic dimension does not arise by adding nodes to a semantic network. Graphical signs have a specified interpretation and serve as means of working with a representation. The second level: the conceptual representation of a particular situation. A situation is cut out from first-level representations as a fragment in light of the position of a participant in joint activity. Its schema-model preserves the organization of that fragment—its elements, structures, connections, processes, functions, evolution, and dynamics. Mel’čuk relates the transition from extralinguistic reality to linguistic sense to a representation of the world as perceived by the speaker.58 The third level: a semantic description of the situation in narrative form. The story elaborates the selected description sequentially through the community’s natural language. A narrative elaborates content sequentially through linguistic schemas; a graphical schema allows representations of processes, structures, and their connections to be maintained and compared jointly and simultaneously. The description of participants, connections, properties, and conditions must preserve the content of the situation. A story differs from what needs to be conveyed in a particular communication. The fourth level: the full sense of the message. Communicative goals guide the selection and organization of content. The full sense represents the content of a particular message about a situation; the speaker’s intention is not included in it. Essential conditions, negation, and what remains unknown are maintained here. Conveying everything necessary for a particular message does not mean telling everything about the situation. This level is a representation of sense, not the text that expresses it. The fifth level: text, the linguistic realization of the message. The language model selects words, syntax, and the order of exposition. Abbreviation is permissible if the recipient can reconstruct the full sense at the fourth level through jointly constructed and agreed-upon representations, the situation of activity, and previous communication. It must not introduce new assertions or change connections, conditions, or what remains unknown. In the reverse direction, the full sense of the message is reconstructed first, 58. I. Mel’čuk, Vers une linguistique Sens-Texte [Toward a Meaning–Text Linguistics] (Collège de France, 1997) (inaugural lecture), https://olst.ling.umontreal.ca/pdf/Mel %27c%CC%8Cuk_1997_Coll%C3%A8ge_de_France.pdf.

21

followed by an extended description of the situation, then its schema-model and the systemic foundations from which it was cut out. Reconstruction does not mean reconstructing the complete original situation from a limited message: the work concerns the connections essential to understanding, not every detail of the situation. Connections that are neither expressed in the message nor confirmed in joint activity remain unknown and must not be filled in as established facts. The extended description makes it possible to identify which connections are necessary for understanding and which still require clarification before being incorporated into the situation schema and related to systemic foundations. The two directions of work across the five levels are shown in the diagram (Figure 1).

Figure 1. Five Levels of Content Organization in ASI Communication

The incompleteness of a systemic representation is consistent with A. Korzybski’s structural differential: a perceived object is an abstraction from a process, and its verbal description is a further abstraction. At each transition, some characteristics remain outside the resulting representation. Korzybski emphasizes that “the object is not the event […] and that the label is not the object nor the

22

event.” A description must therefore be distinguished from what it describes.59 G. P. Shchedrovitskiy provides a theoretical foundation for systemic representation. In “Two Concepts of System” (1974), he writes: “To consider an object as a complex system means to represent it successively in four categorial planes: processes of some one kind, functional structure, organized forms of material, morphology…”.60 In the later account, “Methodology and Philosophy of Organizational and Managerial Activity,” G. P. Shchedrovitskiy distinguishes five planes: processes, functional structures, structures of connections, organized forms of material, or morphology, and material. He explains the transition from four to five by separating functional structures from structures of connections.61 This passage is reproduced in A. P. Zinchenko’s reader.62 “Programming Scientific Research and Development” likewise formulates the requirement to stratify an object into “at least five planes of existence”.63 The systemic planes specify the organization of first-level representations. 59. A. Korzybski, Science and Sanity: An Introduction to Non-Aristotelian Systems and General Semantics at 389 (1933) (ch. XXV: On the Structural Differential; pp. 386– 411; quotation at 389), http://esgs.free.fr/uk/art/sands-ch25.pdf. 60. G. P. Shchedrovitskiy, Два понятия системы [Two Concepts of System], in Избранные труды [Selected Writings] (1974) (electronic publication), https: //gtmarket.ru/library/basis/3961/3972. 61. G. P. Shchedrovitskiy, Методология и философия организационноуправленческой деятельности [Methodology and Philosophy of Organizational and Managerial Activity] (From the Archive of G. P. Shchedrovitskiy, vol. 5, 2003) (quotation republication), https://priss-laboratory.net.ru/S.C.H.E.M.E.S.-/museum -schemes/museum-schemes_2.htm. 62. A. P. Zinchenko, Путеводитель по основным понятиям и схемам методологии Организации, Руководства и Управления [A Guide to the Basic Concepts and Schemas of the Methodology of Organization, Leadership, and Management], at PDF p. 76 (Delo, Moscow, 2004) (G. P. Shchedrovitskiy excerpt at PDF p. 76), https: //pqm-online.com/assets/files/lib/books/zinchenko.pdf. 63. G. P. Shchedrovitskiy, Программирование научных исследований и разработок [Programming Scientific Research and Development] §2.7.2 (From the Archive of G. P. Shchedrovitskiy, vol. 1, 1999) (pt. I, ch. 2, §2.7.2), https://gtmarket.ru/library/basi s/6473/6476.

23

Examining foundations means establishing which schemas and systemic representations have been adopted and which connections have been accepted as true in forming a message. Adjustment means explicitly changing them, including revising distinctions and connections. This work differs from selecting additional texts for the context. Scientific facts and definitions are not ready-made givens independent of ways of constructing the object of inquiry, conducting research, and securing collective recognition. A representation of a situation must therefore distinguish what is established from what is assumed and what remains unknown, and disclose the grounds of that distinction. Ludwik Fleck’s concept of a thought style addresses the historical and collective conditioning of facts and definitions.64 Mel’čuk’s full semantic representation, SemR, provides support for the fourth level. In his 2014 work, Mel’čuk describes the SemR of a sentence as a combination of semantic, semantic-communicative, rhetorical, and referential structures.65 An earlier work already considers the SemR of an utterance realized by a sequence of sentences and the construction of semantic subnetworks for individual sentences.66 The present architecture uses this four-component characterization in organizing the sense of a message in joint activity and defines its place in relation to narrative and the recipient’s understanding. Moreover, words do not have definitive meanings established independently of human history and activity. Language is not only used to express content that has already been constructed; it is also continually transformed in the development of new representations. On the question of the definiteness of linguistic meaning, this paper adopts a polemical position toward Mel’čuk. Kulagina and Mel’čuk characterize two expressions as having “the same and quite definite meaning,” which can refer to 64. W. Sady, Ludwik Fleck, Stanford Encyclopedia of Philosophy (substantive revision: Dec. 18, 2025) (§§ 5–6: the formation of facts and definitions within a thought style), https://plato.stanford.edu/entries/fleck/. 65. I. Mel’čuk, Dependency in Language, in Dependency Linguistics: Recent Advances in Linguistic Theory Using Dependency Structures at 4–5 (K. Gerdes, E. Hajičová & L. Wanner eds., John Benjamins, 2014) (components of the full SemR at 4–5), https://olst.ling.umontreal.ca/pdf/Mel%27c%CC%8Cuk_2014_Dependency_in_La nguage.pdf. 66. Mel’čuk, Meaning-Text Models, supra note 18, at 34, 44.

24

different situations.67 The author rejects treating the definiteness of meaning as an initial characteristic of an expression: in the proposed formulation, its foundation must be established through acquired representations and the practice of use in a particular situation of activity. Stable usage is possible, but it is subject to revision when these foundations change. The very organization of a picture of the world undergoes historical transformation, not merely the stock of information within it. The foundations for identifying objects, determining their properties, and establishing connections change. The persistence of individual terms and results does not attest to the invariance of the whole. An explicit systemic foundation is therefore not a definitive universal ontology: it must be available for examination and revision. In P. G. Shchedrovitskiy’s text “The Idea of ‘Thought=Communication’: Sense vs. Content,” schemas are considered means of organizing thinking: The systemic thought-activity approach assumes that these units are “schemas,” and that the key process is “schematization”.68 The norms of language and thinking are explicitly distinguished: The fact that the rules (norms) of “logic” and “grammar,” and, more broadly, the norms of “language” and the norms of “thinking,” do not coincide was discovered a very long time ago.69 The source itself limits the application of the idea of thoughtcommunication: Thought=communication takes shape only within complete systems of thought-activity.70 67. Kulagina & Mel’čuk, Automatic Translation, supra note 16, at 142. 68. Shchedrovitskiy, Идея «мысли=коммуникации» [The Idea of “Thought=Communication”], supra note 24, §20. 69. Shchedrovitskiy, Идея «мысли=коммуникации» [The Idea of “Thought=Communication”], supra note 24, §18. 70. Shchedrovitskiy, Идея «мысли=коммуникации» [The Idea of “Thought=Communication”], supra note 24, §23.

25

The disagreement with Chomsky also concerns the role of language in constructing thought. Chomsky and his coauthors associate the operation Merge with the construction of hierarchically organized objects of the “language of thought.”71 In the proposed architecture, content organization is not derived from linguistic computation: cutting out a situation and forming a message rely on systemic representations of the world of activity and connections accepted as true. The question is therefore not only how an expression is constructed, but also on what foundations the processes, structures, and connections described are identified. Language’s participation in the development of representations is acknowledged, but this does not make the linguistic mechanism a sufficient foundation of content. Mel’čuk and Polguère write: “we are looking for the semantic primes of individual languages and not for semantic primes of ‘human thought.’ ”72 The systemic categories of the present architecture organize representations of the world of activity, whereas linguistic semantic units organize the sense of its description. Systemic representations are acquired and transformed in communication connected with activity. A participant cuts out a particular situation from these representations and determines the content of a message in joint work. This is the common point of disagreement with the positions of Mel’čuk and Chomsky considered here. Chomsky’s description of an internal linguistic mechanism and Mel’čuk’s account of the transition from a situation through conceptual representation to sense and text do not disclose this organization of the acquisition and use of systemic representations. At the same time, different scientific schools may divide the same domain of reality differently and construct different objects of inquiry within it. Comparing schools requires reconstructing the accepted principles of identifying objects and connections, including distinctions for which another system has no direct equivalent. Kulagina and Mel’čuk distinguish a common 71. J. J. Bolhuis et al., How Could Language Have Evolved?, 12 PLOS Biology e1001934 (2014) (no. 8; coauthors: I. Tattersall, N. Chomsky, R. C. Berwick), https://doi.org/ 10.1371/journal.pbio.1001934. 72. I. Mel’čuk & A. Polguère, Theory and Practice of Lexicographic Definition, 19 Journal of Cognitive Science 417, 460–461 (2018) (no. 4; pp. 417–470), https://olst.ling.um ontreal.ca/pdf/Mel%27c%CC%8Cuk_Polgu%C3%A8re_2018.pdf.

26

situation being described from identity of linguistic meaning.73 The present architecture locates the comparison of ways of constructing an object of inquiry in work on systemic foundations, not merely on linguistic semantics. On the one hand, translation is impossible as a guaranteed substitution of words that preserves content independently of the foundations of the original message. Its task is to reconstruct the essential content of the original description, the expression of which in another language may require creating new linguistic means. On the other hand, translation is possible only insofar as participants can establish correspondences between working representations in a particular situation of activity. Coordinating the limited working ontology necessary for joint activity, rather than an entire picture of the world, is sufficient for this purpose. A shared vocabulary serves as a means of such coordination and is developed or refined in its course. Kulagina and Mel’čuk provide situational support for this proposition. Considering expressions with different linguistic meanings, they write: “they turned out to be translation (not meaning!) equivalents owing only to the fact that their different meanings can describe the same situation.”74 3. Architectural Implications

The proposed organization shifts control of a linguistic response from the finished text to transformations of content. For each transition, the input information, permissible operations, and criteria for preserving content must be specified. Missing information may be obtained from a new source, inferred through an explicitly stated rule, or clarified with the recipient. Changing a foundation differs from clarifying a situation and changing a message. Revising an acquired schema, a systemic distinction, or a connection accepted as true requires checking whether the situation cut out and the content based on it remain valid. Clarifying information or the boundaries of a situation may require revising the narrative and the sense of the message without changing the first level. A change in the communicative goal changes the selection of content, whereas reformulation concerns its linguistic realization. In each case, the representations affected and the consequences of the change must be established 73. Kulagina & Mel’čuk, Automatic Translation, supra note 16. 74. Kulagina & Mel’čuk, Automatic Translation, supra note 16, at 145.

27

in both directions: for the formation of a description and message, and for the reverse reconstruction of the situation and its systemic foundations. Control checks whether a process has been replaced by its result, whether structure, action, and function have been conflated, and whether roles have been attributed to participants or structures without grounds. Control operates both in the movement toward text and in the movement from text to foundations. In the forward direction, it checks the preservation of accepted content during elaboration, selection, and linguistic realization. In the reverse direction, it checks the foundations for reconstructing the full sense, extended description, and situation schema, and identifies the systemic distinctions used. A discrepancy may require clarification of the text or sense, revision of the description and of how the situation is cut out, or a return to systemic foundations. The location at which an error is detected does not limit the direction in which it can be corrected. Separating the levels makes it possible to localize errors: a categorial error at the first level, cutting out a wrong fragment or element at the second, distortion of the story at the third, erroneous selection or organization of a message at the fourth, and ambiguous formulation or unjustified abbreviation at the fifth. Weaver distinguishes the technical accuracy of symbol transmission, the accuracy of conveying the intended sense, and the effect of the received sense on the recipient’s behavior.75 These questions concern different aspects of a single act of communication. Evaluating language-based ASI requires distinguishing the correctness of message reproduction, the reconstruction of its content and foundations, and the use of what has been understood in joint activity. The preservation of a message’s sense must be considered in light of the position of the person understanding it. The source describes this as follows: In developed forms, highlighting such relations of sense presupposes the “speaker’s” adoption of the position of the “person understanding,” simulation of the processes of “understanding” and “non75. W. Weaver, Recent Contributions to the Mathematical Theory of Communication, in C. E. Shannon & W. Weaver, The Mathematical Theory of Communication at 4 (1949) (p. 4; scan of reprint, tenth printing, 1964), https://monoskop.org/images/b/be/Sha nnon_Claude_E_Weaver_Warren_The_Mathematical_Theory_of_Communicati on_1963.pdf.

28

understanding,” and the introduction into the initial text of additional constructions oriented toward dialogue.76 3.1. Understanding a Message and the Criterion of ASI Adequacy

Language and understanding are different processes: knowing words does not guarantee understanding, and the absence of a shared language does not make understanding impossible. P. G. Shchedrovitskiy shows that verbal meanings alone are insufficient for the formation of sense: “But imagine that introducing meaning does not help: all the words have been explained, yet it is still not clear; no sense is formed. Then another function must be brought into play: knowledge must be drawn upon”.77 This distinction also matters for evaluating language AI: explaining the words used does not yet attest to reconstruction of the situation and understanding of the message. In a limited situation, participants’ actions may serve as means of establishing sense without a verbal message. Discussing an ambiguous text, Kulagina and Mel’čuk write that “we cannot rely on our knowledge of the respective language” and that “what we need is knowledge of corresponding reality.”78 P. G. Shchedrovitskiy distinguishes the positions of the speaker and the person understanding.79 A discrepancy requires establishing which representations must be revised and how this will change the content of the message. Recognizing words and the verbal form of a request does not mean under76. Shchedrovitskiy, Идея «мысли=коммуникации» [The Idea of “Thought=Communication”], supra note 24, §21. 77. P. G. Shchedrovitskiy, Выготский Л. С. и современная педагогическая антропология [L. S. Vygotsky and Contemporary Pedagogical Anthropology], in Тьюторство: идея и идеология : материалы 1-й Межрегиональной тьюторской конференции, Томск, 29 февраля — 2 марта 1996 г. [Tutoring: Idea and Ideology: Proceedings of the First Interregional Tutoring Conference, Tomsk, February 29–March 2, 1996] 73 (I. D. Proskurovskaya ed., School of Cultural Policy, Tomsk, 1996) (pp. 73–128; third lecture, section «Знаки, смыслы, значения и знания как организованности мыследеятельности» [“Signs, Senses, Meanings, and Knowledge as Organized Forms of Thought-Activity”]), https://shchedrovitskiy.com/vygotskij-i -sovremennaja-pedagogicheskaja-antropologija. 8 7 . Kulagina & Mel’čuk, Automatic Translation, supra note 16, at 144. 79. Shchedrovitskiy, Идея «мысли=коммуникации» [The Idea of “Thought=Communication”], supra note 24, §4.

29

standing the task. Without a representation of activity, a participant may consider the message understandable yet incorrectly identify the required action. Maturity here is defined by mastery of the means of organizing thinking: participants construct systemic representations of activity and use schemas to coordinate what must be communicated and understood. The completeness of such coordination concerns the essential content of the message, not all the speaker’s individual experiences. In organizing collective thinking, the author observes a distinction between constructing a representation of activity and producing expected utterances about it. A participant may command the terminology, reproduce accepted explanations, and consistently connect phrases that fit one another while failing to maintain within a representation the distinctions between processes and structures and the connections of the activity described. In such cases, verbal coherence does not express constructed subject-matter content but substitutes for the work of constructing it. The continuation of speech is determined by which phrase fits the preceding one and is expected by other participants, rather than by which established connection must be expressed. When expected phrases are being produced, the organization of the activity described is not presented for examination and verification. A clarifying question may elicit a new verbal explanation that maintains the coherence of the conversation without resolving the discrepancy discovered in the organization of the object of inquiry. The mere fact that a conversation is maintained does not mean that participants have common representations of activity. Their joint construction is a separate undertaking, not an automatic result of exchanging messages. The requirement to represent processes and structures graphically reveals the difference between these ways of working. A participant must separate process from structure, show connections between the organized forms identified, relate the sequence of events to the simultaneously existing organization of activity, and preserve accepted distinctions when considering different fragments. What is unknown must be marked as unknown, rather than concealed by a familiar formulation. As participants jointly construct and agree upon a schema, different representations of the organization of activity are presented together, and the discrepancies discovered become objects of discussion and transformation of representations. In communication connected with activity, participants acquire and transform their own and others’ representations, as well as jointly constructed and agreed-upon representations through questions, comparison of explanations, identification of discrepancies, and modification of schemas. 30

Speech substitutes for the construction of content when further speaking is directed toward maintaining the persuasiveness of an answer and a contradiction discovered is covered over with another expected phrase. The observed similarity between this kind of human speaking and languagemodel generation concerns the way the work is performed. Producing an expected verbal continuation substitutes for constructing and checking the systemic representations that provide foundations for both an utterance about activity and the organization of the activity itself. Reproducing ordinary human speech therefore does not provide a sufficient criterion for strong language AI. The organization of content formation must be deliberately developed in both humans and AI: establishing the foundations for cutting out a situation, constructing representations, checking their transformations, and using the results of those checks when forming a message. The criterion is the ability to move from what has been said to a representation of the object of inquiry, work with it, and change the message after revising that representation. It is precisely this ability that makes it possible to distinguish verbal success from technologically organized thinking in joint activity. Boroday maintains that “any correlation of a sign with subject-matter reality takes place through the prism of mental content; that is, reference is impossible without conceptualization”.80 In this formulation, the foundation of the correlation is located in the individual psyche: a sign is correlated with the content of the individual’s conceptual system, through which subject-matter reality is recognized. Although Boroday takes into account the social origins of linguistic conventions and the influence of cultural practices, the immediate explanation of reference proceeds through individual mental content. In the author’s view, the foundation for correlating a sign with reality should be sought in a situation of activity. A participant uses a sign while occupying a particular position in joint work and correlates it with what has been identified as the object of that work. Establishing what an utterance refers to therefore means reconstructing the situation within which it is produced and used. A participant’s mental content may accompany this process, but it does not, in itself, determine which correlation is accepted in the activity or on what founda80. S. Yu. Boroday, Язык и познание: Введение в пострелятивизм [Language and Cognition: An Introduction to Postrelativism] at 583 (V. A. Lektorsky & A. V. Smirnov eds., Sadra LLC: YaSK Publishing House, Moscow, 2020) (800 pp.; ISBN 978-5906859-62-4; propositions cited at 583–584, 593).

31

tions other participants can understand and check it. A graphical schema serves here as a means of bringing a representation into joint work: participants use it to relate what has been said to the organization of activity and to discover discrepancies that cannot be established merely by comparing verbal formulations. For language AI, this distinction determines the object of development. Correlating words with other words or with a model’s latent features does not yet disclose their connection with a situation of activity. The machine must be able to reconstruct and construct a representation of that situation, cut out from systemic representations of activity, and to form a message on its basis. What is then checked is the preservation in the utterance of the processes, structures, and connections described. It is precisely this organization that makes the foundations of a message available for joint examination and transformation. Boroday considers meaning in use to result from its construction in working memory: “Semantics receives its dynamic embodiment in working memory, where meaning is constructed. Such construction takes place on a verbal basis”.81 He relates the selection of activated content to context: “Which particular information enters working memory upon activation depends on a number of factors: linguistic, physical, and social context, background knowledge, the characteristics of communication in a given community, and others”.82 At the same time, “long-term memory stores and continually updates an array of diverse and poorly structured information, while working memory realizes the dynamic interaction of linguistic form, sensorimotor systems, and numerous cognitive operations”.83 Boroday separates successful communication from completeness of understanding: “The answer is simple: successful communication does not require exact translation and complete understanding”.84 Discussing differences between the speaker’s and listener’s representations, he specifies: “Does this mean that the listener will not fully understand the speaker? If understanding means complete reconstruction of the mental model, then yes. However, it does not follow that communication has not occurred”.85 He describes the foundation of such interaction as follows: “If we analyze typical communicative situations even within a 81. Boroday, Язык и познание [Language and Cognition], supra note 80, at 583. 82. Boroday, Язык и познание [Language and Cognition], supra note 80, at 583. 83. Boroday, Язык и познание [Language and Cognition], supra note 80, at 584. 84. Boroday, Язык и познание [Language and Cognition], supra note 80, at 593. 85. Boroday, Язык и познание [Language and Cognition], supra note 80, at 593.

32

single language, we will see that they unfold and are maintained on the basis of core information, whereas the periphery remains underspecified and therefore becomes the prerogative of individual imagination”.86 In the author’s view, these propositions describe immature communication and the work of an immature, non-technologized intellect, for which the foundations of understanding remain implicit and the coordination of content depends on participants’ situational associations. Reproducing ordinary human speech is insufficient if the task is to ensure technologically organized strong thinking. A machine may form expected phrases from context while a person supplies their content from their own representations. Communication will occur, but the work of reconstructing the situation and maintaining its organization will remain with the person. Situational verbal success in such cases does not attest to the machine’s ability to work with the representations on whose basis a message is formed and understood. Strong AI must be able to construct and transform systemic representations of activity while preserving within those representations the distinctions between processes and structures and the connections between them. The similarity of its responses to immature, non-technologized human communication therefore does not justify the machine’s limitations; rather, it shows which way of working the machine reproduces. Context dependence and underspecification observed in ordinary human speech cannot, in themselves, serve as a norm for such AI. Its adequacy must be judged by its ability to perform organized work with the representations necessary for forming a message and understanding it in joint activity. 3 . 2 . Te s t i n g t h e A r c h i t e c t u r a l P r o p o s a l

The next research step is to implement a limited subject-domain example of joint work with AI trained to reconstruct the five levels of communicative content, and to compare response generation using it with ordinary generation. Testing must separately assess the establishment of the accepted foundation and the consequences of changing it, reconstruction of the schema-model and the ability to work with it, preservation of the situation’s content in the message, and the recipient’s understanding. What matters is the preservation of participants, negation, conditions, uncertainty, cost, and temporal distinctions, as well as the consistency of the components of the full sense. 86. Boroday, Язык и познание [Language and Cognition], supra note 80, at 593.

33

The proposed experimental test uses existing tasks with underspecified conditions. Ambig-SWE and Dialogue SWE-Bench make it possible to investigate how a language agent identifies missing information, obtains it through dialogue with a user, and uses it to solve a programming task.8788 Both approaches use SWE-bench Verified material,89 which is based on real SWE-bench tasks.90 QuestBench makes it possible to test separately whether an agent can identify a missing condition and select a question to establish it.91 This material can be used to compare ordinary response generation with the five-level organization of content by evaluating reconstruction of the situation and the message’s sense, the quality of clarifications, and task solution. Authorial corrections, the use of corrected distinctions in new tasks, and the revision of representations when 87. S. Vijayvargiya et al., Ambig-SWE: Interactive Agents to Overcome Underspecificity in Software Engineering, arXiv:2502.13069 (2025) (coauthors: X. Zhou, A. Yerukola, M. Sap, G. Neubig), https://arxiv.org/abs/2502.13069. 88. B. King & J. Flanigan, Dialogue SWE-Bench: A Benchmark for Dialogue-Driven Coding Agents, arXiv:2606.13995 (2026), https://arxiv.org/abs/2606.13995. 89. N. Chowdhury et al., Introducing SWE-bench Verified, OpenAI (2024), https://open ai.com/index/introducing-swe-bench-verified/. 90. C. E. Jimenez et al., SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, ICLR, arXiv:2310.06770 (2024) (coauthors: J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, K. Narasimhan), https://arxiv.org/abs/2310.06770. 91. B. Z. Li et al., QuestBench: Can LLMs ask the right question to acquire information in reasoning tasks?, NeurIPS, arXiv:2503.22674 (2025) (coauthors: B. Kim, Z. Wang), https://arxiv.org/abs/2503.22674.

34

processes, structures, and their connections change are tracked separately. Conclusion

The five-level communication architecture for strong language-based ASI makes the foundations of an utterance explicit and available for examination and adjustment, instead of using so-called “context” and “weights” as the foundation of a response. It connects acquired schemas and systemic connections with cutting out a situation, message formation, and understanding, and makes it possible to trace the consequences of changing a foundation. The proposal also includes reverse work: reconstructing, from a message, the situation and the systemic foundations from which it was cut out, while checking the connections preserved, lost, and changed. The next step for LLMs is the reconstruction of large systemic models of activity, as distinct from successful text and code generation.

35

D E C L A R AT I O N S

Disclosure of AI Assistance AI-assisted tools were used for language editing, structural review, and citation audit support during manuscript preparation. The author remains solely responsible for the conceptual analysis, source selection, citation verification, and final text. Declaration of Competing Interests The author declares that they have no competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Author's Statement on Defensive Publication The author publishes the five-level architecture for language AI described in this article as a defensive disclosure intended to establish a public prior-art record and prevent subsequent patenting of the disclosed combination. Acknowledgments I am grateful to my mother, Valentina Osovskaya, a linguist and a student of Alexey Fyodorovich Losev (1893–1988)—the philosopher, philologist, and scholar of antiquity who wrote on the philosophy of language, the name, and the symbol—for the love of words, language scholarship, linguistics, and the history of languages that she kindled in me from childhood. 36

Related documents

• From Deliberately Constructed Systemic Representations of the World of Activity to Dialogue: A Five-Level Communication Architecture for Strong Language-Based ASI#1448384
• Lost with a Map: Conversational State and Behavioral Reliability in Language Models#1118956
• Talking Machines: Philosophical Essays on Large Language Models#80553
• Does Linguistic Structure Enrichment Enhance Coherence Assessment? Not With Current Architectures#998468
• Percepciones de estudiantes EFL sobre ChatGPT para el desarrollo de la destreza oral#1394446