This paper presents Project Aletheia, a systematic 83-phase investigation of LLM hallucination through the lens of condensed matter physics. Using GPT-2 (124M) as a "particle accelerator, " I establish seven fundamental laws and four theorems governing how transformers suppress factual knowledge in favor of grammatical fluency. Degeneracy Law: Fact-skill angular separation = 1. 2° Temperature Irrelevance: Critical spike is T-independent (γ = 0. 000) LayerNorm Impermeability: All mid-layer interventions absorbed Truth Scaling Law: spikec ~ N−0. 491 Temporal Persistence: Half-life = 130. 9 tokens Grammatical Suppression of Facts (GSF): 70% of facts suppressed by final layers; L9H6 (+927) and L11H7 (+816) identified as primary "Grammar Police" heads Code Mode Switch (NEW): Any non-natural-language symbol prefix (#, //, --) triggers a mode transition that increases suppressor entropy and reduces GSF to near-zero Key New Results in v3 Grammar Police Hierarchy (P61): L11H7 is the top suppressor head (+829 rank degradation) ; fact-grammar crossover occurs at exactly Layer 11 GSF Is Fact-Specific (P63): Math tokens experience zero suppression (0. 4 vs 9. 9 for facts) Dark Matter Hypothesis (P75): Math tokens interact 29% less with suppressor weight matrices, explaining their immunity Code Mode Switch (P76–P80): Symbol-prefixed prompts achieve 25% accuracy vs 0% natural language; all symbols equally effective — a general mode transition, not symbol-specific Full 144-Head Suppression Map (P83): True top suppressor is L9H6 (+927), not L11H7; peak suppression layer is L9, not L11; Head 7 is a suppressor across multiple layers Internal Impossibility extended (P71–P72): Prompt engineering and combined pipelines also fail — baseline outperforms all "improved" methods Phases 1–58 from v2 fully preserved Acknowledgments This research was conducted entirely independently, without institutional affiliation or corporate funding. The author currently faces financial constraints that make it increasingly difficult to maintain subscriptions to AI services essential for this line of research. To sustain and improve the quality of future work, the author is actively seeking community sponsorship. Details are available at https: //github. com/sponsors/hafufu-stack.
Hiroto Funasaki (Wed,) studied this question.