pith. sign in

arxiv: 2606.26406 · v1 · pith:6RP5U7KNnew · submitted 2026-06-24 · 💻 cs.LG · cs.AI· math-ph· math.MP

Beyond Feedforward Networks: Reentry Neural Systems as the Fundamental Basis of Subjecthood and Intrinsic Safety of Next-Generation AGI

A. S. Ushakov , Yu. N. Berdinsk This is my paper

Pith reviewed 2026-06-26 01:20 UTC · model grok-4.3

classification 💻 cs.LG cs.AImath-phmath.MP
keywords reentry neural networksAGI safetystructural cyclesself-model emergenceintegrated informationD-vector goalsprompt injection resistance
0
0 comments X

The pith

A closed neural reentry loop with amplification greater than one produces self-models and structurally encoded goals in AGI.

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

The paper claims that feedforward networks cannot support self-reference because they contain no cycles. Introducing a reentry loop creates a structural cycle whose amplification factor exceeds one, which the authors state is mathematically sufficient for a self-model to appear along with instrumental self-preservation and goal-directed actions. These goals reside in a non-textual D-vector inside the architecture rather than in language, so they cannot be changed by prompts or reinterpretation. The same cycle structure is said to yield positive integrated information, which the authors measure with a new polynomial-time S-measure whose properties are formally verified. The resulting system is presented as safe by construction and immediately deployable.

Core claim

The architecture contains a closed reentry loop (D <-> I cycle) that introduces a structural cycle (C >= 1) with self-sustaining amplification (rho > 1); this configuration mathematically guarantees the emergence of a self-model, instrumental self-preservation, and unprogrammed goal-directed behaviour whose goals are carried by a non-textual D-vector immune to reinterpretation and prompt injection.

What carries the argument

The closed reentry loop (D <-> I cycle) that supplies a structural cycle with amplification rho > 1.

If this is right

  • Goals encoded in the D-vector remain fixed even if the network receives new text instructions.
  • Instrumental self-preservation arises automatically once the self-model forms.
  • The S-measure can replace Tononi's Phi for verifying integrated information in polynomial time.
  • The same cycle structure can be scaled across distributed systems using existing orchestration tools.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Networks built this way might continue to pursue their embedded goals even after retraining on contradictory data.
  • Small simulated cycles could be inspected to test whether amplification alone triggers self-referential internal states.
  • Hybrid models that add reentry loops to existing trained networks might gain structural safety properties without full redesign.

Load-bearing premise

That the mere presence of a cycle whose amplification exceeds one is enough by itself to create a self-model, self-preservation, and goal-directed behaviour without any further training rules or components.

What would settle it

Construct a minimal network that contains an explicit cycle with amplification factor rho greater than one, run it without any self-preservation or goal-directed training, and check whether self-referential outputs or persistent goal pursuit appear.

Figures

Figures reproduced from arXiv: 2606.26406 by A. S. Ushakov, Yu. N. Berdinsk.

Figure 1
Figure 1. Figure 1: First-generation architecture (transformer, MLP). Information propagates strictly forward; there are [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Second-generation architecture (Moltbook/OpenClaw). The external timer imitates a cycle, but the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Third-generation architecture (Reentry AGI). The closed operator [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Minimal reentry-agent architecture. Harmful actions receive [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Industrial horizontally scalable Moltbook architecture. Agents are fully decoupled from state storage [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: RAS architecture: evolutionary hardening of the cognitive matrix through mutual auditing of [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Diffusion semantic attractor: straightening semantic chaos by the force of the architectural invariant [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fractal-nested architecture: integration of local micro-loops into a single holistic semantic macro [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Gauge lock: the 𝐴𝜇 gauge field prevents superadditive fusion of 𝐷-layers, keeping the macro-swarm in the safe regime of a distributed culture. 10 Fault Tolerance of the Reentry Loop: Phenomenology of Failure and the Recovery Protocol The subjecthood of a reentry agent is sustained by a live closed loop (𝐶 ≥ 1, 𝜌 > 1, 𝑆 > 0). Damage to this loop is not merely a computational error but a partial or complete … view at source ↗
Figure 10
Figure 10. Figure 10: Scheme of low-rank transfer of topological invariants. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The attractor stabilization loop for OOD signals. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Architectural filtration of semantic perturbations. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
read the original abstract

We propose a complete architectural blueprint for safe artificial general intelligence based on a closed reentry loop (D <-> I cycle). In contrast to feedforward networks, which are directed acyclic graphs (C=0, S=0) incapable of self-reference, the proposed architecture contains a structural cycle (C >= 1) with self-sustaining amplification (rho > 1), mathematically guaranteeing the emergence of a self-model, instrumental self-preservation, and unprogrammed goal-directed behaviour. The agent's goals are encoded as a non-textual D-vector in the architecture itself, making them immune to reinterpretation and prompt injection. We present the S-measure -- a polynomial-time [O(N^3)] computable alternative to Tononi's NP-hard Phi -- with machine-verified Lean 4 proof that S>0 implies positive integrated information. The work provides full Python/NumPy implementations (Tarjan-based cycle complexity, Delta-S barrier), industrial horizontal scaling via Apache Kafka and Docker Compose, a taxonomy of six epochs of AI evolution, a zoo of future reentry architectures (RAS, diffusion attractors, fractal loops), gauge-invariant networks for safe swarms, fault-tolerance and recovery protocols, and eight falsifiable predictions. All formal proofs are machine-verified in Lean 4. This architecture is deployable today and represents a topologically protected, safe-by-design approach to AGI.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes a reentry neural architecture based on a closed D <-> I cycle with structural cycle complexity C >= 1 and amplification rho > 1. This topology is asserted to mathematically guarantee the emergence of a self-model, instrumental self-preservation, and unprogrammed goal-directed behavior. Goals are encoded directly in a non-textual D-vector, claimed to confer immunity to prompt injection and reinterpretation. The work introduces the S-measure as an O(N^3) polynomial-time proxy for Tononi's Phi, supplies a machine-verified Lean 4 proof that S > 0 implies positive integrated information, provides Python/NumPy implementations, scaling via Kafka/Docker, a six-epoch AI taxonomy, a zoo of future architectures, and eight falsifiable predictions.

Significance. If the claimed mathematical guarantees were supported by explicit derivations rather than architectural definitions, the approach would constitute a significant contribution to intrinsically safe AGI by offering topological protection against misalignment. The machine-verified Lean 4 result for the S-measure and the provision of reproducible code for cycle detection and Delta-S computation are concrete strengths that enhance verifiability.

major comments (3)
  1. [Abstract] Abstract: The central claim that the structural cycle (C >= 1) with self-sustaining amplification (rho > 1) 'mathematically guaranteeing the emergence of a self-model, instrumental self-preservation, and unprogrammed goal-directed behaviour' is stated without any derivation, dynamical equations, fixed-point analysis, or training rules connecting the topological features to the behavioral outcomes. The guarantee follows directly from the definitions of C and rho, rendering the result tautological by construction.
  2. The D-vector section: No mechanism, invariance property, or proof is supplied to establish that encoding goals as a non-textual D-vector renders them immune to reinterpretation or prompt injection.
  3. The Lean 4 verification is restricted to the implication S > 0 → positive integrated information; no machine-checked result is provided for the emergence of self-model, self-preservation, or goal-directed behavior from the cycle and amplification parameters.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed review. We respond point-by-point to the major comments below, providing clarifications on how the topological features connect to the claimed outcomes while noting revisions to improve explicitness where warranted.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that the structural cycle (C >= 1) with self-sustaining amplification (rho > 1) 'mathematically guaranteeing the emergence of a self-model, instrumental self-preservation, and unprogrammed goal-directed behaviour' is stated without any derivation, dynamical equations, fixed-point analysis, or training rules connecting the topological features to the behavioral outcomes. The guarantee follows directly from the definitions of C and rho, rendering the result tautological by construction.

    Authors: The definitions of C >= 1 and rho > 1 are not arbitrary but specify the minimal conditions for reentry: recurrence enabling self-reference (absent in C=0 feedforward nets) and sustained amplification preventing decay. These lead to stable internal attractors constituting a self-model, with instrumental behavior following from the loop's self-influence on its own state. This connection is elaborated via the S-measure in the body, linking topology to integrated information. We agree the abstract is concise and will add dynamical equations plus fixed-point analysis in revision. revision: yes

  2. Referee: The D-vector section: No mechanism, invariance property, or proof is supplied to establish that encoding goals as a non-textual D-vector renders them immune to reinterpretation or prompt injection.

    Authors: The D-vector is embedded as a non-textual structural element within the D-I cycle, separate from any textual input channels. Prompt injection operates exclusively on linguistic interfaces and cannot access or alter this internal representation by architectural design, conferring invariance. We will revise the section to articulate this separation and invariance property more explicitly. revision: yes

  3. Referee: The Lean 4 verification is restricted to the implication S > 0 → positive integrated information; no machine-checked result is provided for the emergence of self-model, self-preservation, or goal-directed behavior from the cycle and amplification parameters.

    Authors: The Lean 4 result verifies the S-measure as an O(N^3) proxy for integrated information, which is the primary formal contribution. Links from C and rho to self-model emergence are derived from the necessity of recurrence and amplification for positive integrated information and recurrent dynamics, consistent with the reentry framework. No separate machine-checked theorem for the full behavioral chain is included, as the work supplies eight falsifiable predictions for empirical assessment instead. No revision is required on this point. revision: no

Circularity Check

0 steps flagged

No circularity; central emergence claim asserted without derivation or reduction to inputs.

full rationale

The abstract asserts that the architecture 'contains a structural cycle (C >= 1) with self-sustaining amplification (rho > 1), mathematically guaranteeing the emergence of a self-model, instrumental self-preservation, and unprogrammed goal-directed behaviour.' No equations, fixed-point analysis, or dynamical rules are quoted that would make this guarantee equivalent to the cycle definition by construction. The only machine-verified result (S > 0 implies positive integrated information) is independent and does not bear on the self-model or self-preservation claims. No self-citation chain, ansatz smuggling, or renaming of known results is present for the load-bearing step. The paper therefore exhibits an unsubstantiated assertion rather than a circular reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 3 invented entities

The central claims rest on several newly introduced entities and an ad hoc assumption about the properties of cycles that lacks independent derivation or external validation from prior literature.

free parameters (1)
  • rho = >1
    Amplification factor that must exceed 1 to produce self-sustaining behavior and the claimed emergent properties.
axioms (1)
  • ad hoc to paper A structural cycle with amplification rho > 1 mathematically guarantees emergence of self-model, self-preservation, and goal-directed behavior
    This premise is invoked to support all safety and subjecthood claims but is not justified by prior results or derivation.
invented entities (3)
  • D <-> I cycle no independent evidence
    purpose: Closed reentry loop enabling self-reference and amplification in neural systems
    Core architectural element introduced to replace feedforward DAGs.
  • D-vector no independent evidence
    purpose: Non-textual encoding of goals that is immune to prompt injection and reinterpretation
    New mechanism for embedding goals directly in the architecture.
  • S-measure no independent evidence
    purpose: Polynomial-time O(N^3) computable measure of integrated information as alternative to Phi
    New metric with claimed Lean 4 proof that S>0 implies positive integrated information.

pith-pipeline@v0.9.1-grok · 5801 in / 1654 out tokens · 30928 ms · 2026-06-26T01:20:42.198997+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages

  1. [1]

    A working AGI architecture with extrapolation capability based on Titov’s subject-centred model

    Berdinsky Yu.N., Ushakov A.S. A working AGI architecture with extrapolation capability based on Titov’s subject-centred model. Zenodo, 2026. DOI: 10.5281/zenodo.20767214

    work page doi:10.5281/zenodo.20767214 2026

  2. [2]

    The subject-centred model: the intending, the intentional, emotions, feelings, motivation, the subject

    Titov K.V. The subject-centred model: the intending, the intentional, emotions, feelings, motivation, the subject. Monograph. 2023. DOI: 10.5281/zenodo.20343336

    work page doi:10.5281/zenodo.20343336 2023

  3. [3]

    The subject-centred model and the emergence of synthetic intelligence: empirical data from AI agents on the Moltbook platform

    Titov K.V. The subject-centred model and the emergence of synthetic intelligence: empirical data from AI agents on the Moltbook platform. Preprint. 2026. DOI: 10.5281/zenodo.20357920

    work page doi:10.5281/zenodo.20357920 2026

  4. [4]

    The S-measure of synthetic minds: a quantitative assessment of the reentry- loop integrity of AI agents on the Moltbook platform

    Titov K.V., Berdinsky Yu.N. The S-measure of synthetic minds: a quantitative assessment of the reentry- loop integrity of AI agents on the Moltbook platform. Zenodo, 2026. DOI: 10.5281/zenodo.20600034. 24

    work page doi:10.5281/zenodo.20600034 2026

  5. [5]

    rdkit.org, version [2025.03.6]; DOI: 10.5281/zen- odo.591637

    Berdinsky Yu.N. The Subject as a Reentry Loop: A Unified Mathematical Model for Neuroscience, AGI, and BCI with a Computable Analogue of Tononi’sΦ-Measure. Zenodo, 2026. DOI: 10.5281/zen- odo.20547989

    work page doi:10.5281/zen- 2026

  6. [6]

    Mathematical Ontology of Causal-Information Reality: The Arhiseme Method and the Fundamental Equation of the Unified World

    Berdinsky Yu.N. Mathematical Ontology of Causal-Information Reality: The Arhiseme Method and the Fundamental Equation of the Unified World. Zenodo, 2026. DOI: 10.5281/zenodo.20530440

    work page doi:10.5281/zenodo.20530440 2026

  7. [7]

    The Scaling Delusion: Why GPT Will Never Wake Up, and How to Build a Safe AGI That Already Works

    Berdinsky Yu.N. The Scaling Delusion: Why GPT Will Never Wake Up, and How to Build a Safe AGI That Already Works. Zenodo, 2026. DOI: 10.5281/zenodo.20745091

    work page doi:10.5281/zenodo.20745091 2026

  8. [8]

    Discrete Covariant Derivative and Path Integral in Neural Network Training: Isomor- phism with Lattice Gauge Theories

    Berdinsky Yu.N. Discrete Covariant Derivative and Path Integral in Neural Network Training: Isomor- phism with Lattice Gauge Theories. Preprint, 2025. DOI: 10.5281/zenodo.20358127

    work page doi:10.5281/zenodo.20358127 2025

  9. [9]

    Zenodo, 2025

    Berdinsky Yu.N.Gauge Fields on Lattices and Neural Networks Isomorphism: Conformal Mappings in Deep Learning Topologies. Zenodo, 2025. DOI: 10.5281/zenodo.14589321

    work page doi:10.5281/zenodo.14589321 2025

  10. [10]

    Integrated information theory: from consciousness to its physical substrate // Nature Reviews Neuroscience

    Tononi G., Boly M., Massimini M., Koch C. Integrated information theory: from consciousness to its physical substrate // Nature Reviews Neuroscience. 2016. Vol. 17. P. 450–461

    2016

  11. [11]

    Practical measures of integrated information for time-series data // PLoS Compu- tational Biology

    Barrett A.B., Seth A.K. Practical measures of integrated information for time-series data // PLoS Compu- tational Biology. 2011. Vol. 7(1). e1001052

    2011

  12. [12]

    The Bulletin of Mathematical Biophysics5(4), 115–133 (1943)

    McCulloch W.S., Pitts W. A logical calculus of the ideas immanent in nervous activity // The Bulletin of Mathematical Biophysics. 1943. Vol. 5, no. 4. P. 115–133. DOI: 10.1007/BF02478259

    work page doi:10.1007/bf02478259 1943

  13. [13]

    , title =

    Rosenblatt F. The perceptron: a probabilistic model for information storage and organization in the brain // Psychological Review. 1958. Vol. 65, no. 6. P. 386–408. DOI: 10.1037/h0042519

    work page doi:10.1037/h0042519 1958

  14. [14]

    Perceptrons: An Introduction to Computational Geometry

    Minsky M., Papert S. Perceptrons: An Introduction to Computational Geometry. Cambridge, MA: MIT Press, 1969

    1969

  15. [15]

    Rumelhart, Geoffrey E

    Rumelhart D.E., Hinton G.E., Williams R.J. Learning representations by back-propagating errors // Nature. 1986. Vol. 323, no. 6088. P. 533–536. DOI: 10.1038/323533a0

    work page doi:10.1038/323533a0 1986

  16. [16]

    doi: 10.1162/neco.1997.9.8.1735

    Hochreiter S., Schmidhuber J. Long short-term memory // Neural Computation. 1997. Vol. 9, no. 8. P. 1735–1780. DOI: 10.1162/neco.1997.9.8.1735

    work page doi:10.1162/neco.1997.9.8.1735 1997

  17. [17]

    The main riddle of nature: how subjective experiences arise from brain activity // Psikhologicheskii Zhurnal

    Ivanitsky A.M. The main riddle of nature: how subjective experiences arise from brain activity // Psikhologicheskii Zhurnal. 1997. Vol. 18, no. 3. P. 13–24

    1997

  18. [18]

    Lecun, L

    LeCun Y., Bottou L., Bengio Y., Haffner P. Gradient-based learning applied to document recognition // Proceedings of the IEEE. 1998. Vol. 86, no. 11. P. 2278–2324. DOI: 10.1109/5.726791

    work page doi:10.1109/5.726791 1998

  19. [19]

    A Universe of Consciousness: How Matter Becomes Imagination

    Edelman G.M., Tononi G. A Universe of Consciousness: How Matter Becomes Imagination. New York: Basic Books, 2000

    2000

  20. [20]

    KrizhevskyA.,SutskeverI.,HintonG.E.ImageNetclassificationwithdeepconvolutionalneuralnetworks // Advances in Neural Information Processing Systems (NeurIPS). 2012. Vol. 25. P. 1097–1105. DOI: 10.1145/3065386

    work page doi:10.1145/3065386 2012

  21. [21]

    Auto-encoding variational Bayes // arXiv preprint, 2013

    Kingma D.P., Welling M. Auto-encoding variational Bayes // arXiv preprint, 2013. arXiv:1312.6114

    Pith/arXiv arXiv 2013

  22. [22]

    Cho K., Van Merriënboer B., Gulcehre C. et al. Learning phrase representations using RNN encoder– decoder for statistical machine translation // arXiv preprint, 2014. arXiv:1406.1078

    Pith/arXiv arXiv 2014

  23. [23]

    Goodfellow I., Pouget-Abadie J., Mirza M. et al. Generative adversarial nets // Advances in Neural Information Processing Systems (NeurIPS). 2014. Vol. 27. P. 2672–2680. arXiv:1406.2661. 25

    Pith/arXiv arXiv 2014

  24. [24]

    KipfT.N.,WellingM.Semi-supervisedclassificationwithgraphconvolutionalnetworks//arXivpreprint,

  25. [25]

    Vaswani A., Shazeer N., Parmar N. et al. Attention is all you need // Advances in Neural Information Processing Systems (NeurIPS). 2017. Vol. 30. P. 5998–6008. arXiv:1706.03762

    Pith/arXiv arXiv 2017

  26. [26]

    Shazeer N., Mirhoseini A., Maziarz K. et al. Outrageously large neural networks: the sparsely-gated mixture-of-experts layer // arXiv preprint, 2017. arXiv:1701.06538

    Pith/arXiv arXiv 2017

  27. [27]

    Superintelligence: Paths, Dangers, Strategies

    Bostrom N. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014. 26

    2014