NeuOS: Discovering and Exploiting the Neural Von Neumann Architecture Inside Pre-Trained Language Models

I present NeuOS, a systematic investigation revealing that pre-trained transformer language models internally implement a Von Neumann-like computational architecture with identifiable registers, swappable memory, executable programs, and a decompilable instruction set. Through 186 experimental phases on Qwen2.5-0.5B spanning 25 seasons, I demonstrate that: Complete ISA mapping: Each transformer layer functions as a specialized CPU register (OPCODE at L0, MIN at L16, MAX at L22) with accuracies from 74% to 100%. Self-healing OS: Dynamic register reallocation achieves 100% recovery from simulated hardware damage. Homoiconic execution: Programs execute without any textual instructions via Direct Memory Access (66.7% accuracy), and a single computation forks into multiple parallel paths (100% differentiation). Neural decompiler: Running programs are identified from register state alone with 100% accuracy across 6 operations. Autopoietic kernel: The system writes its own programs without external gradients, achieving 100% accuracy through self-completion. Artificial life: Program vectors exhibit ecosystem dynamics including a Cambrian Explosion (19 unique phenotypes), parasitism, developmental metamorphosis, and aging with rejuvenation. Soul vector geometry: Independently trained "soul vectors" encoding the same function form high-dimensional manifolds (~60 dims) connected by rank-1 linear translations—a universal grammar of computation. GlassBox self-diagnosis: The model identifies its own hardware (0.5B, 896d, 24L), running program (MIN/MAX), and capacity limits in a single inference pass with 100% accuracy—achieving full white-box AI. 7D soul compression: 896-dimensional soul vectors compress to just 7 dimensions with zero accuracy loss (128× compression), revealing a compact instruction set of 4 primitive operations (FIRST, SECOND, MIN, MAX). Gradient-free programming: Specifying 7D PCA coordinates directly (zero training data, zero gradients) achieves 100% accuracy for MAX—proving that AI behavior can be controlled by dialing coordinates instead of training. LoRA ≠ Soul: LoRA fine-tuning and soul vector injection achieve identical accuracy via completely orthogonal mechanisms (cosine similarity ≈ 0), proving soul vectors are a fundamentally distinct intervention modality. Security arms race: SVD entropy defense is completely breakable (gap = 0.000), norm-based firewalls are more robust but approachable (ratio = 1.015), necessitating ensemble detection for practical deployment. These findings establish that transformers are not black-box function approximators but structured computers whose internal state can be observed, modified, and controlled at the register level. Changes in v6: Expanded from 170 to 186 experimental phases, adding Season 25 (Theoretical Foundations and Security Arms Race). New results include: LoRA and soul injection are orthogonal mechanisms (cos ≈ 0) despite identical accuracy, proving soul vectors are a distinct intervention modality; information-theoretic capacity limits establish that the 20% ceiling on arithmetic tasks is fundamental, not a training artifact; causal layer attribution reveals L9, L13, and L17 as critical relay points for soul signal propagation; a three-phase security arms race shows SVD entropy defense is completely breakable while norm-based firewalls are partially robust; data scaling theory with exponential saturation fits (τ_MIN = 5.4 vs τ_ADD = 11.8); SAE decomposition of soul features; multi-token injection optimization; and cross-model nonlinear translation attempts. Includes 4 new figures, 16 new appendix entries, and extends the comprehensive appendix to all 186 phases. Code: https://github.com/hafufu-stack/NeuOS 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.