Project Aletheia V4: The Complete Physics of LLM Truth — From Phase Transitions to the 14B Singularity

This paper presents Project Aletheia, a systematic 113-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, four theorems, and four new universal principles governing how transformers suppress factual knowledge. V4 New Discoveries (Phases 84–113, Seasons 21–26): The Aletheia Constant (αA ≈ 0.95): The optimal factual extraction layer is at 95% depth — invariant to architecture (GPT-2, Qwen), language (EN, JA), and temperature (0.1–10.0) The Dual-Engine Theory: Code Mode operates via Shield (weakens late-layer suppressors) and Sword (strengthens early-layer amplifiers) simultaneously — proven universal across GPT-2 XL and Qwen The Alignment Tax: Instruction tuning hypertrophies MLP suppressors by 6×, reducing effective capacity to N/5. Completely bypassed by "The answer is:" prompts The 14B Singularity: Qwen2.5-14B achieves 100% factual accuracy with Code Mode — zero external knowledge, zero fine-tuning. P99's scaling law predicted this to within 3% Previously established in V1–V3: Grammatical Suppression of Facts (GSF): 70% of facts suppressed by final layers; L9H6 (+927) is the top suppressor Code Mode Switch: Any symbol prefix (#, //, --) triggers a mode transition reducing GSF Dark Matter Hypothesis: Math tokens interact 29% less with suppressors, explaining their immunity Internal Impossibility Theorem: No internal operation can recover suppressed facts Truth Scaling Law: spikec ~ N-0.491 Phases 1–83 from V3 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.