This paper presents Project Aletheia, a systematic 138-phase investigation of LLM hallucination through the lens of condensed matter physics. Using GPT-2 (124M) as a "particle accelerator" and scaling to Qwen2.5-14B, I establish seven fundamental laws, six theorems, and four principles governing how transformers suppress factual knowledge. V6 New Discoveries (Phases 130–138c, Seasons 29–30): Embedding Dispersion Surgery: Physically separating clustered numerical embeddings (cosine 0.73 → 0.05) enables DPO to achieve 50% numerical accuracy — up from 0%, causally proving that embedding geometry is the source of numerical immunity. The L2 Distance Law: DPO's critical condition is not cosine similarity but L2 distance. The Gram-Schmidt Paradox shows that perfect orthogonality (cos = 0) fails; only L2 > 1.2 enables DPO. This is the third phase transition discovered in the project. Shield and Sword Protocol: Combining Code Mode (shield against suppressors) with FGA injection (sword for fact promotion) achieves 43% numerical accuracy with zero training — purely inference-time intervention. The Ultimate Combo: Unifying Surgery + DPO + Shield L9H6 (+927) is the top suppressor Code Mode Switch: Any symbol prefix (# // -- *) triggers a mode transition reducing GSF The 14B Singularity: Qwen2.5-14B achieves 100% factual accuracy with Code Mode DPO Suppression Theorem: DPO suppresses rejected tokens (100% reliability), not promoting correct ones (73%) Aletheia Constant (αA ≈ 0.95): Universal across architecture, language, and temperature Phases 1–129 from V1–V5 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 (Fri,) studied this question.