The Experiential–Factual Divide in LLM Hidden-State Geometry: Six Converging Lines of Evidence Across 7–8 Architectures

This paper presents six converging lines of evidence that the Experiential–Factual (E–F) geometric divide — previously identified in static word embeddings across seven languages by Alieksieienko (2026) — constitutes a universal structural constraint on large language model (LLM) hidden-state representations. Across 7–8 architectures spanning 2019–2024 (GPT-2-XL, OPT-6.7B, Falcon-7B, Mistral-7B, Gemma-2-9B, Llama-3.1-8B, Qwen2.5-7B, Phi-3-mini), we demonstrate: (1) Robust E–F geometric separation in all architectures (t = 17.4–23.2, all p < 0.0001)(2) Cross-architecture causal transfer — GPT-2-XL (2019) predicts errors in 7 modern models (mean Spearman ρ = 0.912, permutation p = 0.0009)(3) Semantic alignment with Tulving's episodic/semantic memory distinction (Cohen's d = 0.70, p < 0.0001)(4) Higher output uncertainty for Experiential questions (t = 2.07, p = 0.039)(5) Universal symmetric attractor structure (r = 0.998, p < 0.0001, across 7 architectures)(6) Causal E-axis activation steering selectively modulates Experiential content (ΔE = −0.179, p = 0.050), confirmed against random-vector ablation control These findings converge with Anthropic's April 2026 discovery of functional emotion vectors (Transformer Circuits Thread) and the common low-dimensional semantic subspace reported by Schiekiera et al. (2026). The E–F divide represents a previously uncharacterised universal constraint: language models are systematically less reliable precisely where human cognition relies most heavily on experiential, cultural, and subjective processing. Research conducted in collaboration with Claude (Anthropic).Full replication code and data included.