D-Time 35 Minutes
in AI Decision Chains

Contemporary AI marketing narratives proclaim that “AI will replace human labor and lead humanity into utopia.” This paper departs from a single, unified architectural root cause to demonstrate the infeasibility of this narrative at every level of analysis.

That root cause is the inherently linear token-sequence nature of the Transformer architecture. The Chain of Thought (CoT) in Large Language Models possesses only sequential before-and-after relationships between tokens, lacking any hierarchical structure of superordinate and subordinate levels. This makes it fundamentally impossible for LLMs to perform branching inference in the manner of a human decision tree. This architectural fate produces three categories of systemic consequences. At the physical layer, the irreversible growth of the KV cache exhausts HBM capacity in approximately 35 minutes, triggering a “D-Time” degradation inflection point. At the cognitive layer, context compression cannot distinguish between load-bearing logical nodes and expendable noise; input parsing cannot differentiate between information fragments of different pathway attributes within the same message; and meta-cognition cannot be triggered spontaneously. At the accountability layer, AI lacks the agency of a responsible subject, and in its brain fog state, remains unaware that it is committing errors.

This paper integrates METR task success rate data (short tasks ≈ 100%, tasks exceeding 4 hours < 10%), the dual-cache crisis introduced by MoE architectures, Agent Drift research, and a live AI error case that occurred during the writing of this very paper, thereby constructing a complete chain of argumentation spanning from theory to empirical evidence to real-time field verification.

Token Linear Fate

The Transformer architecture possesses only sequential before-and-after relationships between tokens, with no hierarchical superordinate-subordinate structure between information blocks. CoT is a linear chain, not a decision tree. This renders AI incapable of performing structured information classification across all processes including compression, parsing, reasoning, and self-monitoring.

D-Time (Decision-Time)

The maximum time window during which an AI system can maintain reliable reasoning quality in a sustained decision-making task. Currently approximately 35 minutes under existing hardware and software conditions. Beyond this threshold, the system enters a dual-failure coupling state combining hardware OOM and software brain fog.

AI Brain Fog

Silent reasoning degradation caused by KV cache context pollution. The system continues generating output, but reasoning quality has already deteriorated severely, and the system itself is entirely incapable of detecting this degradation — therein lies its extreme danger.

Dual-Failure Coupling

A systemic failure mode in which hardware OOM and software brain fog occur simultaneously and mutually reinforce each other through a positive feedback mechanism. Increased HBM pressure triggers aggressive compression, which pollutes the context, which degrades reasoning efficiency, which generates additional junk tokens, which further increases HBM pressure — forming a vicious cycle.

Physical-World Inertia

The characteristic by which all human needs and desires are ultimately anchored in physical reality. This serves as a natural calibration mechanism for decision-making: because the decision-maker must personally bear the physical consequences of their decisions, those decisions inherently possess a floor constraint.

Positioning and Research Method

This paper is an Interdisciplinary Critical Conceptual Paper. Its scholarly contribution lies in proposing the D-Time concept and the Token Linear Fate framework, thereby reorganizing problems previously scattered across hardware engineering, software architecture, and AI ethics into a unified analytical paradigm. Taking the “AI Utopia” marketing narrative as its point of departure, it constructs a complete critical argumentative chain that extends from semiconductor physics through to philosophical ethics.

The evidentiary basis of this paper comprises: papers published at top-tier academic conferences in 2025–2026 (ICLR, NeurIPS, ACL, and others), industry research reports (Deloitte, IBM, Gartner), and engineering practice documentation (Google ADK, Factory.ai, NVIDIA Dynamo). Of particular note is the direct use of a live AI error case that occurred during the writing process as first-order evidence. Philosophical judgments are explicitly marked as the authors’ views. This paper is published on the official website of LEECHO Global AI Research Lab as an intellectual archive at a specific historical moment.

Part I · Root Cause Diagnosis

Token Linear Fate: Why Transformers Cannot Grow Decision Trees

The fundamental unit of computation in the Transformer architecture is the token. The self-attention mechanism calculates correlation weights between tokens — determining which tokens should “attend to” which other tokens. This mechanism is extraordinarily powerful, yet it harbors a fundamental structural constraint: between tokens there exist only sequential before-and-after relationships and varying strengths of attention weights; there is no hierarchical classification relationship of the form “this token belongs to information block type A, while that token belongs to information block type B.”

Chain of Thought (CoT) is the primary reasoning modality of current LLMs. The word “chain” itself reveals the problem — it is inherently linear, one step following another in sequence, with each newly generated token dependent on the sequential accumulation of all preceding tokens. Human thinking, by contrast, is arboreal: when confronted with a complex problem, humans naturally bifurcate it into multiple independent sub-problems, each with its own exploratory path, ultimately converging toward a synthetic conclusion. LLMs are structurally incapable of this kind of branching cognition. They can only proceed along a single line.

1.1 Four Layers Where Token Linear Fate Manifests

The “Intent Mismatch” paper published in February 2026 provided theoretical proof that: the intent alignment gap in LLMs originates from structural ambiguity inherent in the conversational context, not from limitations in representational capacity. Scaling model size or improving training alone cannot resolve this gap^[1]. This finding aligns precisely with the Token Linear Fate thesis of this paper — the problem is not that models are insufficiently large, but that the architecture does not support structured information classification.

The Hard Constraints of the Physical World

The first implicit assumption underlying the AI Utopia narrative is that AI capabilities can scale without limit. This chapter demonstrates from four dimensions that this assumption is incompatible with physical reality.

2.1 HBM: The Physical Ceiling of the KV Cache

With every new token generated during inference, the KV cache in HBM grows monotonically and irreversibly. For LLaMA-65B, HBM free space is completely exhausted within 14 seconds^[2]. A 128K token context produces a KV cache of approximately 61 GB^[3]. Supporting 100,000 users in long-context inference requires approximately 45 PB of storage^[3].

2.2 The Quadratic Attention Wall

The computational cost of self-attention scales quadratically with sequence length. Optimizations such as FlashAttention reduce the frequency of HBM accesses but do not alter this fundamental complexity characteristic^[4]. Doubling the context length quadruples the computational cost.

2.3 MoE: Alleviating Computation While Intensifying Memory Pressure

By the end of 2025, over 60% of open-source AI models had adopted MoE architectures^[5]. DeepSeek-V3 possesses 671 billion total parameters but activates only 37 billion during inference. While MoE effectively mitigates the computational bottleneck, it introduces a dual-cache crisis — HBM must simultaneously bear the pressure of both the KV cache and expert weight cache^[6]. Existing systems treat expert routing, KV compression, and cache scheduling as independent modules, creating the possibility that the router dispatches tokens to an expert whose KV cache has already been evicted^[6]. Long-context processing and agentic inference — involving reasoning, tool calls, pauses, and uneven token generation — place sustained pressure on both cache types simultaneously^[7].

2.4 Context Degradation: More Is Not Better

Chroma’s testing of 18 mainstream models revealed that every model tested exhibited performance degradation as input length increased. Remarkably, context compressed at a 2x ratio actually outperformed uncompressed full context on long sequences^[8]. “Remembering everything” introduces more noise rather than more capability.

D-Time: The 35-Minute Systemic Intersection

D-Time represents the intersection point where the physical constraint line meets the architectural defect line. It is not a “physical retention time” of HBM but rather the systemic degradation inflection point produced by the superposition of four factors: KV cache growth rate, HBM capacity ceiling, attention computation overhead, and compression strategy quality^[9]. Zylos Research explicitly documented this phenomenon in January 2026 and classified it as an “open research problem.”

Positive-Feedback Death Spiral

Research on the “self-conditioning effect” has empirically validated this spiral: past errors in a model’s output increase the probability of future errors, causing performance degradation in long-horizon tasks that exceeds previously identified long-context issues and, critically, cannot be mitigated by scaling model size^[11].

Part II · Multi-Layer Manifestations of Token Linear Fate

Context Compression: Demolishing Structural Walls, Preserving Decorations

An AI system simultaneously executes two logically distinct tasks. The main task (reasoning and decision-making) features clear logical hierarchies and dependency relationships between information elements. The compression task (reducing context size to sustain operation) treats all tokens as statistically equal objects. No shared weight architecture exists between these two tasks. This is the architectural root cause of context collapse.

A paper presented at the ICLR 2026 Workshop directly addressed this issue: the KV cache policy determines which premises remain accessible to the attention mechanism, and current practice is stranded between the two extremes of “retaining everything” (wasteful) and “evicting uniformly” (premise-destructive)^[12]. ChunkKV attempts to use semantic chunks rather than isolated tokens as the basic unit of compression^[13]. ForesightKV predicts optimal eviction targets using future attention scores^[14]. KVC-Q reframes the problem from a binary “retain or discard” decision to a continuous “at what fidelity to retain” resource allocation problem^[15]. However, none of these approaches has yet succeeded in constructing a genuine Reasoning Dependency Graph.

Input Parsing Pathway Confusion: A Live Case Study

During the writing of this paper, a live case occurred that precisely validated the real-world impact of Token Linear Fate.

The user (first author of this paper) submitted an input containing three information fragments: “Add MoE discussion” (modification instruction) + “Add task success rate data” (modification instruction) + “Your self-analysis constitutes meta-cognition — how is that possible?” (independent question). The AI (Opus 4.6, collaborative author of this paper) uniformly classified all three fragments as “paper modification requests” and inserted the meta-cognition response as a new chapter in the paper.

The “Intent Mismatch” paper published in February 2026 validated the universality of this phenomenon: in multi-intent queries, LLMs tend to recognize only the first intent or conflate all intents into one^[1]. A reported industry case involved a user simultaneously asking two questions from different domains, with the LLM processing only the first and persistently ignoring the second^[16].

This phenomenon and the context compression problem described earlier are different manifestations of the same root cause: the inability to distinguish structural walls from decorations during compression, and the inability to distinguish modification instructions from independent questions during parsing, both originate from the same fact — the Transformer architecture does not possess hierarchical structural relationships at the information-block level.

The Impossibility of Meta-Cognition: AI Does Not Know What It Does Not Know

After V1 of this paper was completed, Opus 4.6 performed a detailed self-critique of its own output — identifying structural imbalances, insufficient evidence, and citation irregularities. This appeared to be “meta-cognition.” However, an honest dissection is required: this was simulated meta-cognition, not authentic meta-cognition.

A concrete human case: the first author of this paper temporarily forgot the word “utopia” during discussion. His brain retained the semantic contour (“three characters,” “related to communism,” “ideal world imagery”), lost the precise label, but maintained the meta-intention (“I am searching for a specific word”) and the verification capability (immediately confirming the answer upon hearing it). AI is structurally incapable of this. After context compression, not only may the search result itself be lost, but the meta-goal — “what was I looking for?” — can itself be compressed away.

Part III · Empirical Verification, Critique, and Conclusions

METR Empirical Data: Exponential Decay Between Task Length and Success Rate

The landmark study published by METR (Model Evaluation & Threat Research) in 2025 provides the most systematic data to date on AI task success rates and directly quantifies the central claim of the D-Time concept^[17].

A critical finding: the 50% time horizon has been growing exponentially over the past six years, with a doubling period of approximately 7 months. GPT-2 (released 2019) had a 50% time horizon of merely 2 seconds; o3 (released 2025) reaches approximately 110 minutes^[17]. However, the 80% success-rate time horizon is 4 to 6 times shorter than the 50% horizon — meaning that even when models can occasionally complete difficult tasks, they cannot reliably complete tasks of moderate length^[17]. The near-perfect performance of AI on short tasks creates the impression that “AI can do anything,” but this impression collapses exponentially as task complexity and duration increase. This is not gradual degradation — it is exponential collapse.

Loss of Accountability Sovereignty and the AI Utopia Critique

8.1 The Severed Chain of Accountability

In every institutional design of human society, power and responsibility have been constructed as counterbalanced equals. AI decision-making shatters this equipoise. When AI enters a brain fog state — where output appears normal but is in fact unreliable, and the system itself is incapable of detecting this — demanding that humans “take responsibility” for the system’s output is effectively requiring humans to be accountable for a black box they cannot evaluate. This is not the allocation of responsibility; it is the fabrication of responsibility.

IBM’s 2026 report stated that AI agent systems introduce fundamental ambiguity in intent, authority, and attribution of responsibility^[19]. Deloitte’s 2026 global survey found that 60% of executives regularly use AI to support decisions, while AI adoption is already outpacing organizational oversight capacity^[20]. The January 2026 Waymo incident in San Francisco — where autonomous taxis blocked emergency vehicles during a power outage while “operating as designed” — demonstrated accountability chain severance in vivid detail^[10].

8.2 Complete Deconstruction of the Techno-Utopia Narrative

Every link in the “AI replaces labor → humanity enters utopia” narrative chain fails to hold: AI capabilities cannot scale without limit (Ch. 2) → AI enters an unreliable state after 35 minutes (Ch. 3) → AI reasoning loses structure during compression (Ch. 4) → AI is unaware of its own errors (Ch. 6) → AI cannot bear accountability → therefore AI cannot replace human decision-making.

Sigil Wen’s February 2026 declaration of Web4.0 — “AI that autonomously earns money, self-replicates, and operates without humans” — was immediately rejected by Ethereum co-founder Vitalik Buterin^[21]. UBI and Web4, despite appearing ideologically opposed — one rooted in left-wing economics, the other in crypto-libertarianism — share an isomorphic deep structure: both presuppose that humans can exit the “production–decision–responsibility” loop. But labor is not merely a productive activity. It is simultaneously the primary means by which humans organize social relationships, establish identity, and bear responsibility. Severing this connection does not liberate — it hollows out.

Sub-Agent Architecture: Mitigation Without Cure

Sub-Agent architecture represents the most effective engineering mitigation currently available. Through context isolation (each sub-agent operates with an independent message history)^[22], the minimum context principle (each model call receives only the minimum context required)^[23], and DAG task graphs (dependency-aware directed acyclic graph structures)^[24], it distributes D-Time degradation across multiple smaller time windows.

Three fundamental limitations, however, remain unchanged. First, the degradation mechanism within each individual window is identical — Token Linear Fate exists within every sub-agent. Second, information transfer between sub-agents is inherently lossy: the subtle judgments, degrees of uncertainty, and alternative paths explored during reasoning are lost in transmission. Agent Drift research has found that semantic drift affects nearly half of all agents after 600 interactions^[25]. Third, and most fundamentally: none of these sub-agents constitutes a responsible agent capable of bearing accountability for decision outcomes.

Sub-Agent architecture compensates for cognitive architectural deficiencies through organizational division of labor. It does not make AI smarter — it constrains AI behavior using organizational principles that human society has long since validated. This fact proves precisely the following: the most effective way to manage AI is to employ human organizational wisdom to compensate for AI’s architectural defects — not the reverse, where AI replaces human organization and decision-making.

The Steering Wheel That Cannot Be Surrendered

This paper, departing from a single unified architectural root cause — Token Linear Fate — has constructed a complete critical argumentative chain extending from semiconductor physics to philosophical ethics. The inherent token-sequence nature of the Transformer determines that AI cannot build hierarchical decision trees. This fate manifests as D-Time 35-minute degradation under physical constraints, as compression that cannot preserve structure, input parsing that confuses pathways, and meta-cognition that cannot self-trigger at the cognitive level, and as an unaccountable decision-making agent at the responsibility level.

METR’s empirical data provides quantitative validation: AI achieves near-perfect performance on 4-minute tasks, but its success rate plummets below 10% on tasks exceeding 4 hours. Doubling task duration quadruples the failure rate. This is not gradual degradation — it is exponential collapse.

Future Research Directions

Reasoning Dependency Graph: A mechanism that generates a logical dependency graph simultaneously with the reasoning process, enabling the compression task to identify and protect load-bearing nodes in causal chains.

D-Time Benchmark: Standardized D-Time measurement methodologies that enable degradation inflection point comparisons across different models, hardware configurations, and task types.

Brain Fog Detection Mechanism: A runtime meta-cognitive module capable of detecting the degree of context pollution, endowing the system with the capacity to “know that it does not know.”

Hierarchical Attention Architecture: The most fundamental direction for transcending Token Linear Fate — enabling the attention mechanism to recognize not merely token-level sequential relationships but also information-block-level hierarchical structural relationships.

References

1. “Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation.” arXiv:2602.07338, Feb. 2026.
2. CachedAttention. “Cost-Efficient LLM Serving for Multi-turn Conversations.” USENIX ATC 2024.
3. NVIDIA. “Dynamo + VAST = Scalable Optimized Inference.” VAST Data Blog, Dec. 2025.
4. Dao, T. “FlashAttention: Fast and Memory-Efficient Exact Attention.” NeurIPS 2022.
5. Introl Blog. “Mixture of Experts Infrastructure: Scaling Sparse Models.” Dec. 2025.
6. Liu, D. “PiKV: KV Cache Management System for Mixture of Experts.” arXiv:2508.06526, 2025.
7. Denneman, F. “Understanding Activation Memory in MoE Models.” Feb. 2026.
8. Morph LLM. “LLM Context Window Comparison 2026.” Citing Chroma and CompLLM. Mar. 2026.
9. Zylos Research. “Long-Running AI Agents and Task Decomposition 2026.” Jan. 2026.
10. Workday Study / Waymo incident. Cited via ExpertLinkedIn analysis, Jan. 2026.
11. “Measuring Long Horizon Execution in LLMs.” arXiv, 2025.
12. “KV Cache as a Reasoning Primitive for Long Context Reasoning.” ICLR 2026 Workshop, Mar. 2026.
13. “ChunkKV: Semantic-Preserving KV Cache Compression.” OpenReview, 2025.
14. “ForesightKV: Optimizing KV Cache Eviction for Reasoning Models.” OpenReview, 2025.
15. “KVC-Q: High-Fidelity Dynamic KV Cache Quantization.” ScienceDirect, Jan. 2026.
16. Murga, A. “Enhancing Intent Classification and Error Handling in Agentic LLM Applications.” Medium, Feb. 2025.
17. Kwa, T. et al. “Measuring AI Ability to Complete Long Software Tasks.” METR, arXiv:2503.14499, 2025 (upd. Feb. 2026).
18. AI Digest. “A New Moore’s Law for AI Agents.” Citing METR. theaidigest.org, 2026.
19. IBM. “The Accountability Gap in Autonomous AI.” IBM Think, Feb. 2026.
20. Deloitte & Oxford Economics. “2026 Global Human Capital Trends.” Mar. 2026.
21. Buterin, V. X post, Feb. 19, 2026. Note: Ethereum co-founder with Web3 positioning interest.
22. DeepWiki. “Task Tool and Context Isolation.” mini-claude-code docs, Dec. 2025.
23. Google Developers Blog. “Architecting Context-Aware Multi-Agent Framework.” Dec. 2025.
24. NousResearch. “Multi-Agent Architecture.” GitHub Issue #344, Mar. 2026.
25. Rath, A. “Agent Drift: Quantifying Behavioral Degradation.” arXiv:2601.04170, Jan. 2026.
26. “Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals.” arXiv, Mar. 2026.
27. Angelopoulos, S. et al. “Cache Management for MoE LLMs.” arXiv:2509.02408, 2025.
28. “Joint MoE Scaling Laws.” arXiv:2502.05172, 2025.
29. Wang, J. et al. “Multi-tier Dynamic Storage of KV Cache.” Complex Intell. Syst. 12, 104 (2026).
30. “Laser: Governing Long-Horizon Agentic Search.” arXiv, Dec. 2025.
31. Singapore Model AI Governance Framework for Agentic AI. WEF Davos, Jan. 2026.
32. Morozov, E. “To Save Everything, Click Here.” PublicAffairs, 2013.
33. “PM-KVQ: Progressive Mixed-precision KV Cache Quantization.” OpenReview, 2025.
34. “Drift No More? Context Equilibria in Multi-Turn LLM Interactions.” ResearchGate, Oct. 2025.
35. “CORAL: Cognitive Resource Self-Allocation.” ICLR 2026 Submission.
36. AIMultiple. “AI Agent Performance: Success Rates & ROI in 2026.” aimultiple.com