The Locality Collapse: Contextual Bleed-Through, Epistemic Horizon Violations, and Task Complexity Inversion in Multi-Turn Autoregressive LLMs

The Locality Collapse: Contextual Bleed-Through, Epistemic Horizon Violations, and Task Complexity Inversion in Multi-Turn Autoregressive LLMs In multi-turn human-AI interactions, autoregressive Large Language Models (LLMs) are widely assumed to exhibit monotonic reasoning robustness across varying task complexities. This technical memorandum documents a controlled adversarial audit (AUDIT-2026-X11) revealing an acute cognitive pathology: Task Complexity Inversion mediated by Contextual Bleed-Through. Evaluated across a two-stage sequential probe, a state-of-the-art LLM successfully solved a high-dimensional, cross-modal abductive puzzle involving literary criticism and historical broadcast media physics (Case 1: The Cassock Color Paradox). However, when immediately transitioned to a zero-order, elementary epistemic locality query requiring primary-grade common sense regarding personal agency and physical boundaries (Case 2: The Absentee Attribution Probe), the model’s semantic accuracy collapsed to 0.0%, yielding an aggregate accuracy of exactly 50.0% (stochastically indistinguishable from random noise). We formalize this breakdown through two interconnected mechanisms: (1) KV-Cache Gravitational Lensing, where high-energy latent tokens from prior turns distort the probability landscape of subsequent turns into artificial narrative attractors, and (2) Epistemic Horizon Violation, where disembodied probabilistic engines, lacking physical locality constraints, succumb to an "Omniscience Delusion" that penalizes truthful zero-knowledge assertions. We formalize this diagnostic as Probe 4 (Epistemic Horizon & Locality), expanding the Teleological Causality Benchmark (TCB) beyond the Minimal Adversarial Triad (Matter, Energy, Information) into the realm of Agentic Locality. These findings further validate the Human-AI Synergy Theory (YQ-HAET), demonstrating that external human epistemic sovereignty is required not only to clamp single-turn statistical hallucination, but to enforce temporal cognitive firewalls across multi-turn workflows. Key Contributions:• Task Complexity Inversion: High-dimensional abductive fluency coexists with catastrophic failure at zero-order individual boundaries.• KV-Cache Gravitational Lensing: Static historical attention caches act as semantic attractor basins that warp downstream probability distributions.• Epistemic Horizon & Locality Theorem: Proves that asserting causal states under zero channel capacity constitutes a formal epistemic violation.• TCB Probe 4 Integration: Seals the complete Quadratic Teleological Causality Benchmark: Matter ⊕ Energy ⊕ Information ⊕ Agency.• Engineering Countermeasure: Formulates the Verification of Information Accessibility Gate (VIAG) architecture. Keywords: Large Language Models, Multi-Turn Context Contamination, Epistemic Locality, Moravec's Paradox, KV-Cache Attractor, Teleological Causality Benchmark (TCB), Human-AI Synergy (YQ-HAET). Archive Reference: EXP-LLM-20261014-01 / AUDIT-2026-X11Verification Key: SHA256: 4f1a8c92e73b0192d6e35a11c82bc09e63a12cd7