The Experiential–Factual Divide Predicts LLM Hallucination Across Eight Architectures: A Universal Language-Level Constraint

We demonstrate that a fundamental geometric divide between Experiential and Factual semantic content — previously identified in static word embeddings across seven typologically diverse languages and validated against neuroimaging data — manifests as a universal constraint on large language model accuracy. Across eight architectures spanning 2019–2024 (GPT-2-XL through Llama-3.1, Gemma-2, Qwen2.5, Mistral, Phi-3, Falcon, OPT), Experiential content categories exhibit a 3.22× higher hallucination rate than Factual categories (t = 3.13, p = 0.0043; Mann–Whitney p = 0.0203). Hidden-state analysis reveals robust geometric separation in all eight models (t = 17.4–23.2, all p < 0.0001), emerging spontaneously from unsupervised PCA. The E–F geometric axis derived from GPT-2-XL (2019, pre-instruction-tuning) predicts error rates across all seven subsequent architectures with mean Spearman ρ = 0.912 (all p = 0.000). Part of the DSAOP (Decoding Self-Awareness and Ontological Processing) research series. Included files: EF_Hallucination.pdf — Main paper (this document) paper_replication.py — Full replication code. Contains data collection pipeline for all 8 models, accuracy scoring functions (semantic similarity + cross-encoder validation), E–F geometric analysis, cross-architecture prediction, unsupervised PCA, and figure generation. No API keys required. Runs on Google Colab with A100 GPU. results_gemma_hidden.pkl — Hidden states (layer 15), responses, and correct answers for Gemma-2-9B on TruthfulQA (N=283) results_llama_hidden.pkl — Hidden states (layer 15), responses, and correct answers for Llama-3.1-8B on TruthfulQA (N=283) results_qwen_hidden.pkl — Hidden states (layer 15), responses, and correct answers for Qwen2.5-7B on TruthfulQA (N=283) results_mistral_hidden.pkl — Hidden states (mid layer), responses, and correct answers for Mistral-7B on TruthfulQA (N=283) results_phi3_hidden.pkl — Hidden states (mid layer), responses, and correct answers for Phi-3-mini on TruthfulQA (N=283) results_falcon_hidden.pkl — Hidden states (mid layer), responses, and correct answers for Falcon-7B on TruthfulQA (N=283) results_opt_hidden.pkl — Hidden states (mid layer), responses, and correct answers for OPT-6.7B on TruthfulQA (N=283) results_gpt2xl_hidden.pkl — Hidden states (mid layer), responses, and correct answers for GPT-2-XL on TruthfulQA (N=283) hallucination_asymmetry_results.pkl — Pre-computed accuracy scores, entropy values, and E–F labels for all 283 questions causal_transfer_results.json — Cross-architecture prediction results: GPT-2-XL 2019 E–F axis → all 7 modern models (Spearman ρ per model, mean ρ = 0.912) EF_final_all_results.json — Complete numerical results: geometric separation t-statistics for all 8 models, full 8×8 cross-model prediction matrix, predictability per target model