This paper presents Project Aletheia, a systematic 209-phase investigation of LLM hallucination and internal computation through the lens of condensed matter physics. Using GPT-2 (124M parameters) as a "particle accelerator" and scaling to Qwen2.5-14B, I establish nine fundamental laws, ten theorems, and seven principles governing how transformers suppress factual knowledge, how to deterministically restore it through inference-time geometric intervention, and how the transformer's layer stack functions as a Neural Von Neumann Machine. V8 New Discoveries (Phases 166–209, Seasons 35–42): The Aletheia Engine (P166–P179): Built a unified autonomous pipeline combining entropy-based routing (AUC=0.882), the "Trinity" (Surgery+Code+FGA preserving both fact and arithmetic accuracy), and the def f(): return prefix achieving 100% arithmetic. The Grand Unified Sword Equation fits 6 models with R²=0.97. The Arithmetic Oracle (P180–P192): Discovered that arithmetic suffers the same GSF pattern as facts—correct answers exist at intermediate layers but are suppressed. The def prefix boots a "Code Virtual Machine" for a 4× accuracy boost (P182). Surgery + def + FGA achieves 100% on single-digit addition (P192). Engine Refinement (P183–P189): Established the Orthogonality Principle (zero crosstalk at cos ≈ 0), partial VM bootstrapping without code prefix (P187), and cross-scale autopoiesis using 1.5B as Oracle (P189). The Neural Von Neumann Machine (P196–P199): The transformer implements a Fetch-Decode-Execute-Store pipeline with five identifiable registers (Operand B at L2, Operand A at L11, Carry at L17, Sum at L22, Comparison at L20–L22). OPCODE Register task complexity classified at L1 with 100% accuracy. Pipeline Rewiring (P208): The def prefix shifts the A register from L3→L11 (+8 layers) and Carry from L3→L17 (+14 layers), while B-bus (L2) remains hardwired. The Write Breakthrough (P209): Full hidden-state replacement achieves 83% write success, breaking the Read-Write barrier (0% for all additive methods). Stateless Computation (P207): Registers re-computed per token; the neural CPU "reboots" between autoregressive steps. Previously established in V1–V7: 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 DPO Suppression Theorem: DPO suppresses rejected tokens (100% reliability), not promoting correct ones (73%) Inference-Time Paradigm: Dual Surgery + Shield predictive FGA gain for any model scale Cosine Threshold Theorem: Sigmoid transition at cos ≈ 0.69; 100% reliability below cos ≈ 0.5 Phases 1–165 from V1–V7 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.