SNN-Genesis v15: Stochastic Resonance in LLM Reasoning — Large-Scale Validation, Reasoning Manifold Fine Structure, and Cross-Architecture Divergence

SNN-Genesis v15 provides large-scale validation, maps the fine structure of the reasoning manifold, and discovers that stochastic resonance is architecture-specific — Qwen2.5-7B shows the opposite noise response pattern to Mistral-7B. NEW in v15 (Season 16, Phases 80, 83, 82): Large-Scale Validation (Phase 80): v14's 50% record (N=30) corrects to 40% (fixed_full) and 37% (dual_anneal) at N=100, both with Fisher p < 0.0001. The 40% level is the true asymptotic ceiling. Reasoning Manifold Fine Structure (Phase 83): Seven-band PCA decomposition reveals the reasoning manifold occupies ~64 of 4096 dimensions. PCA bands 1–64 are harmful/neutral; the transition to beneficial noise occurs at ~PC 65. PCA-bottom (43.3%) ≈ random (46.7%). Cross-Architecture Divergence (Phase 82): On Qwen2.5-7B, random noise harms reasoning (30% vs baseline 46.7%) — the exact opposite of Mistral. PCA-top-4 has no effect (46.7% = baseline). Stochastic resonance is architecture-specific, not universal. From v14 (retained): Reasoning Manifold Geometry (Phase 71): PCA-top-4 noise = baseline; reasoning axis in top PCs Orthogonal Complement = Random (Phase 79): projecting out top-4 PCs has zero effect (46.7%) 50% All-Time Record at N=30 (Phase 78): full-rank + Flash Annealing Honest Correction (Phase 77): recovery-only does not replicate at N=100 From v13 (retained): Low-Rank Noise Efficiency (Phase 70b): k=256 matches full-rank, k=4 achieves 5× baseline Semantic Phase Decomposition (Phase 76): Flash Annealing 30.0% > all-on semantic 23.3% Noise Source Invariance (Phase 70c; honest null result) 1/√N dose law, 40% at N=100, Flash Annealing 85 page paper. Full experimental code and data included. Code: https://github.com/hafufu-stack/snn-genesis