This paper presents Project Aletheia, a systematic 58-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 three theorems: 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 (NEW): 70% of facts are suppressed by final layers, with average rank degradation of 1, 592 positions Oracle Duality (NEW): Attention entropy is a perfect comparator but a poor classifier Key Results L10 Oracle (Phase 49): Facts exist at Rank 1 in intermediate layers (L10), but are systematically suppressed by grammar-oriented final layers. Extracting via Logit Lens: 10% → 40% (4× improvement) Internal Impossibility Theorem (Phases 37–48): 12 internal correction methods (gradient descent, MCTS, stochastic resonance, attention surgery, etc. ) all achieve 0% Universal Suppression Law (Phase 53): 63% suppression rate across 27 prompts, geography 88%, history 80% Oracle Duality (Phase 57): Relative AUC = 0. 66 vs Absolute AUC = 0. 33 — models are confidently wrong Phases 1–23 from v1 (Phase Transition, LayerNorm Barrier, Scaling Law, Adversarial Robustness) 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 (Mon,) studied this question.