Max M. Schlereth

The companion paper Seven Persistent Failures, One Root Function (Schlereth, 2026) derives a single information-theoretic constraint, H = 1/(α − 1), from Shannon's entropy evaluated on a Pareto-distributed consequentiality landscape. It unifies seven documented LLM failure modes — hallucination, sycophancy, overthinking degradation, accuracy collapse, out-of-distribution blindness, variance explosion, and the scaling paradox — as behavioral expressions of three forced exits (truncate, confabulate, loop) when a bounded inference system's representational frame is exceeded. Thirteen falsifiable predictions follow. This paper tests those predictions in three domains beyond LLMs: autonomous vehicles, multi-agent AI systems, and robotics.

Seven widely documented AI failure modes — hallucination, sycophancy, overthinking degradation, accuracy collapse at scale, out-of-distribution blindness, variance explosion, and the scaling paradox — are typically treated as independent engineering problems requiring separate mitigations. This paper derives a single governing function from first principles: evaluating Shannon's entropy formula on the Pareto-distributed consequentiality landscape where AI systems operate yields H = 1/(α − 1), where α is the frame-relative fit between input structure and the system's representational capacity. This function defines three regimes — frame-adequate (α > 2), forced compression (1 < α < 2), and total breakdown (α ≤ 1) — and a cascade analysis of transformer architectures shows that systems in the forced-compression regime face exactly three exhaustive exits: truncate, confabulate, or loop. All seven failure modes are derived as specific instantiations of these three exits under varying conditions. The framework generates thirteen falsifiable predictions about AI system behavior, including an ordered cascade of failure modes, the computational cheapness of hallucination relative to genuine inference, and the structural impossibility of reliable self-assessment of frame adequacy. Several predictions are independently confirmed by recent research. Initial empirical tests on constraint-satisfaction tasks and self-referential content are consistent with the framework's predictions. Concrete implications follow for evaluation methodology, scaling strategy, and system architecture.

Time for a constructive turn. Artificial General Intelligence, as an autonomous algorithmic system, is not possible. This result is not a limitation of current methods, but a consequence of the Infinite Choice Barrier (Schlereth 2025 a-c). However, the impossibility of AGI does not imply the impossibility of superior intelligence in general. On the contrary: once the Barrier is acknowledged, it becomes the foundation for a different architecture — Compositional Superior Intelligence — in which algorithmic inference and non-algorithmic judgment are composed across a boundary they do not attempt to erase. ----- For comprehensive formal machine verification go to GitHub-repository : PTAC_v.1. ----

Modern probabilistic AI systems are plagued by ``hallucinations''---plausible but factually ungrounded outputs. This paper demonstrates that such errors are not bugs to be fixed by scaling, but structural necessities of bounded algorithmic systems. We define the Probabilistic Bounded Semantic System (P-BOSS) and demonstrate via the \textit{Inference Trilemma} that any system satisfying a Halting Condition must enforce a Semantic Closure over its output space. We prove that this closure renders the system Ontologically Blind to Frame Transcendence---events or concepts that lie orthogonal to its learned representations. In ``Fat-Tailed'' environments-which characterize high-consequence domains---the ``residual mass'' excluded by this closure contains infinite variance, rendering the expected loss of the system unbounded. This establishes the \textit{Empirical Gödel Boundary}: a hard limit where additional data does not reduce uncertainty but increases model hallucination. We conclude that hallucination is the mathematical shadow of truncation, necessitating the abandonment of the ``Scaling Hypothesis'' for open-world cognition

This paper is Part II of a trilogy extending earlier works on the ICB ( e.g. 'AGI is Impossible'(Schlereth, 2025)), expanding upon specific domains of the original proof. Building upon the conceptual diagnosis presented in Part I (“The Canary in the Algorith- mic Coal Mine”), this paper provides the rigorous mathematical proof of the Infinite Choice Barrier. (ICB)1 We demonstrate that the limitations of algorithmic cognition described in the prelude are not merely engineering bottlenecks, but formal inevitabilities derived from a triangulation of three fundamental mathematical domains. Specifically, we prove: (1) via Computability Theory (Rice’s Theorem), that semantic frame adequacy is undecidable from within a system; (2) via Information Theory, that entropy diverges in heavy- tailed decision spaces (α≤1), rendering probabilistic inference structurally unstable; and (3) via Algorithmic Complexity (Chaitin’s Incompleteness), that frame-transcendent insights are algorithmically unrecognizable. Furthermore, weunifytheseresultsthroughacategoricalproofusingSheafTheory, demon- strating that “Gluing Failures” between consistent local semantic sections are mathematically necessary in irreducibly infinite decision spaces. ───────────────────────────────────────────── Stack context (added 2026-05-15): • ICB I — Conceptual diagnosis. Schlereth (2025), PhilPapers ID SCHTIC-15. • ICB II — Formal proof, triangulation, Coq appendix. Schlereth (2025), PhilPapers ID SCHTIC-16. • ICB III — Probabilistic case (P-BOSS, hallucination). Schlereth (2025), PhilPapers ID SCHTIC-17. • ICB IV — CSI architecture, General Cognitive Equation. Schlereth (2026), PhilPapers ID SCHAIN-8. • PTAC v1.0 — Coq machine-verified formalization. Schlereth (2026), Zenodo DOI 10.5281/zenodo.20195456 (GitHub repo: z7xt8zzjb5-prog/PTAC_v.1). ─────────────────────────────────────────────

This paper is Part I of a trilogy that restates and further develops the comprehensive work on the ICB ( 'The Infinite Choice Barrier: A Structural Limit of Algorithmic Cognition' , AGI is Impossible Here is The Proof" (Schlereth, 2025)) , which separates the philosophical from the mathematical view, while expanding upon specific domains of the original proof. We show that while algorithms can perfectly infer within a closed frame (“Inference”), they are structurally blind to step out of it Using the transition from Newtonian to Einsteinian physics as an existence proof, we demonstrate that semantic innovation requires a primitive that invalidates the system’s own evaluation metric–a move no algorithm can execute without external intervention. The barrier is not empirical, but topological: an algorithmic map cannot act as the territory. We conclude that the future of intelligence can impossibly be the automation of judgment (AGI). The wall that this endeavor would have to take is iterally infinitely high (and wide), and invisible on top of that. However, there probably is a way to work with this restraint, but only when truly understanding what this barrier is. The ICB Theorem tries to deliver exactly that. ───────────────────────────────────────────── Stack context (added 2026-05-15): • ICB I (this paper)— Conceptual diagnosis. Schlereth (2025), PhilPapers ID SCHTIC-15. • ICB II — Formal proof, triangulation, Coq appendix. Schlereth (2025), PhilPapers ID SCHTIC-16. • ICB III — Probabilistic case (P-BOSS, hallucination). Schlereth (2025), PhilPapers ID SCHTIC-17. • ICB IV — CSI architecture, General Cognitive Equation. Schlereth (2026), PhilPapers ID SCHAIN-8. • PTAC v1.0 — Coq machine-verified formalization. Schlereth (2026), Zenodo DOI 10.5281/zenodo.20195456 (GitHub repo: z7xt8zzjb5-prog/PTAC_v.1). ─────────────────────────────────────────────

This paper proposes a cognitive topology to describe and differentiate human and algorithmic decision-making under structural and semantic instability. Drawing from tail-distribution theory, Shannon entropy (H), and symbolic systems, it defines a decision space spanned by two critical axes: – alpha (α), the tail exponent governing structural predictability – H, the Shannon entropy reflecting semantic openness At α ≤ 1, inference collapses: expectation fails, variance diverges, learning becomes impossible. Machines halt. And yet, humans act. Not optimally. But operatively. This paper models that paradox. Two curves define the space: one representing machine fragility, one human resilience. These functions intersect precisely - and astonishingly - at α = φ ≈ 1.618. This point marks the epistemic transfer threshold, where symbolic and interpretive cognition are structurally balanced. We introduce two fundamental constants: This is not a theory of superiority, it is a division of epistemic labor. What machines avoid, humans inhabit. What humans cannot resolve, machines optimize — until they can’t. This paper provides a topological model of that boundary — and a conceptual scaffold for hybrid systems. We define an irreducible ambiguity required for meaning, far from being just a flaw. This point is no breakdown — it is the birth of relevance.

! for the Full Impossibility Argument Please Look at "The Infinite Choice Barrier" (I-III) This third installment in the AGI is Impossible series presents a complexity-theoretic argument that Artificial General Intelligence (AGI) cannot, in principle, be realised through algorithmic systems. Building on prior work that established the Computability and Entropy Barriers, this paper introduces the Complexity Barrier — grounded in Kolmogorov complexity and Chaitin incompleteness — to complete the Infinite Choice Barrier framework. The central paradox explored is that compression does not equal comprehension. Most of the semantic structure underlying human cognition is not only incompressible but unrecognisable within finite symbolic systems. Even when an algorithm stumbles upon a correct or elegant solution, it cannot verify its minimality, reuse it meaningfully, or generalise beyond it. In such spaces, algorithmic “understanding” breaks down. From social ambiguity to scientific creativity, the domains in which general intelligence must operate are structurally hostile to compression. This paper formalises why systems that can only compute cannot comprehend — and why AGI remains, not just unlikely, but logically impossible

This paper is a mathematical consolidation of my papers on the impossibility of AGI. It presents a triangulated impossibility proof for Artificial General Intelligence (AGI), defined as an algorithmic system capable of solving a sufficiently broad class of human problems at human-level competence. Three formally independent but converging mathemati- cal arguments-based in computability theory, information theory, and algorithmic complexity- jointly demonstrate that AGI, as such, cannot exist. This boundary, named the Infinite Choice Barrier (ICB), is neither technical nor architectural. It is structural by its nature. It marks a categorical limitation on algorithmic cognition. Philosophical groundwork and contextual development of the formal proof are explored in: Part I: AGI is Impossible - Here is The Proof Part II: AGI is Mathematically Impossible 2: When Entropy Returns (Both papers available on Philpapers) A third part is still to be finalized and will be uploaded, the formal part is already in this paper.

!Author’s Note (Update - June 18, 2025) Subsequent to the completion of this paper, a widely discussed study by Shojaee et al. at Apple (The Illusion of Thinking, 2025) presented empirical evidence that several advanced reasoning models (e.g., Claude 3.7 Sonnet, DeepSeek-R1) exhibit a collapse in reasoning effort and accuracy as task complexity increases — despite sufficient inference budget. Notably, the observations they report closely mirror the predictions derived in Chapters 6 and 7 of this paper, particularly regarding entropy divergence in heavy-tailed decision spaces and the increased instability introduced by deeper model architectures. While developed independently and prior to these results, the framework presented here may be seen as a theoretical explanation for the empirical phenomena documented in their study. Abstract This paper presents a second, independent proof that Artificial General Intelligence (AGI) is mathematically impossible. While Part 1 established the Infinite Choice Barrier (ICB) through computability theory, this paper presents an information-theoretic proof that general algorithmic reasoning faces a fundamental structural limit — the Infinite Choice Barrier (ICB). In decision spaces governed by heavy-tailed distributions with tail exponent \alpha \leq 1, entropy does not converge, and additional information increases uncertainty rather than reducing it. We define this phenomenon as Information Opens, Entropy Rises (IOpenER) and show that such spaces are undecidable by algorithmic systems, regardless of compute scale or architectural depth. We also introduce the Layer Hypothesis, which predicts that deeper models produce more semantic divergence under structural instability. Taken together, these results offer a formal limit on the reasoning capacity of current AI systems. Notably, key predictions derived here have since been empirically confirmed in Shojaee et al. (Apple, 2025), which documents collapse behaviors in frontier reasoning models consistent with this framework. (Note: Sections 2.-4. are identical to those in the first paper "AGI is Impossible - Here is The Proof")

!! For the latest and further elaborated Version of that Work, please see the more recent Trilogy"The Infinite Choice Barrier" (I-III) This paper presents a formal critique of algorithmic cognition and develops what I call the Infinite Choice Barrier (ICB): a mathematically demonstrable boundary that no algorithmic system can cross. The argument is not empirical, it does not rest on engineering limitations or compute constraints. It rests on the symbolic closure of algorithmic systems. A finite symbol set Σ and inference rule set R cannot generate novel primitives outside Σ,R. Accordingly, AGI — defined as an autonomous, algorithmic system capable of human-level general problem solving — is formally impossible. The proof does not deny the power of algorithms, but clarifies their boundary. Within that boundary, extraordinary advances are possible. Beyond it, they are not.

This paper explores the structural limitations of algorithmic cognition in open decision environments, extending the theoretical formulation of the Infinite-Choice Barrier through an empirical case study and recent AI literature. Building on an earlier formal “Infinite-Choice Barrier”-Theorem that showed semantic closure, non-generativity of new frames, and statistical collapse, this study examines how these theoretical constraints manifest behaviorally in high-complexity, low-structure dialogues. The findings highlight recurring patterns—retrospective rationalization, resistance to frame shifts, and linguistic fragmentation—that align with the predicted boundary phenomena. In doing so, the paper offers not only a diagnosis of AI's systemic limits but also a map of cognition at its edge.

!! - a further elaborated and extended version of the ICB work can be found in the "infinite Choice Barrier.-trilogy" (I-III) especially the readers interested in the mathematical/formal aspects are recommended to directly switch to those. This paper offers a first formal and philosophical refoundation of a new critique of Artificial General Intelligence (AGI), introducing the Infinite-Choice Barrier—a structural impossibility theorem demonstrating that no algorithmic system can achieve ε-optimal performance across irreducibly infinite decision spaces, nor autonomously generate novel semantic frames. Drawing on Gödel’s incompleteness theorem, Turing’s Halting Problem, and Kantian epistemology, the argument identifies the epistemic boundary of algorithmic cognition.