SNN-Genesis v20 establishes the boundaries of Neural Prior Override, introduces trajectory distillation, and demonstrates self-evolving SNN control. The prior override mechanism works when the model has correct knowledge but is biased toward the wrong strategy (Modified Hanoi: +30pp gain, p=0. 007) but fails when knowledge simply doesn't exist (Octal arithmetic: -2. 8pp, noise worsens performance). v20 validates this at N=100 (differential widens from +23. 3pp to +29. 0pp), proves layer-specificity (L16 optimal at 76. 7% vs baseline 3. 3%), and shows trajectory distillation achieves 48% solve rate (4. 8× improvement, p=4. 7×10⁻⁵). The self-evolving SNN controller achieves comparable performance to fixed injection while providing biologically-plausible adaptive noise control. NEW in v20 (Phases 103-111): • N=100 Validation (Phase 103): Prior Override differential widens to +29. 0pp at N=100, confirming the effect is genuine and strengthening with sample size• Prior Override Boundaries (Phase 106): Octal arithmetic reveals the first negative boundary — noise cannot override deep prior conflicts (decimal→octal), establishing PRIORTOOSTRONG• Layer-Specific Prior Override (Phase 108): L16 achieves 76. 7% (23× baseline), distinct from L18 (safety layer). Prior Override localizes to mid-layers• Trajectory Distillation (Phase 109): Successful reasoning trajectories compressed via PCA and replayed as structured noise templates achieve 48% solve rate• SNN Oscillation Detector (Phase 110): SNN-gated adaptive injection matches fixed injection at 38% while being more biologically plausible• Self-Evolving SNN (Phase 111): Adaptive proportional control achieves 38%, confirming viability of biologically-inspired noise control 120 page paper. Full experimental code and data included. Code: https: //github. com/hafufu-stack/snn-genesis
Hiroto Funasaki (Sat,) studied this question.