Internal Noise as an Anti-Collapse Resource: Preliminary Evidence for a Safe Edge Band in a Transformer-Based Dynamical System

Internal noise in neural systems is typically treated as a source of performancedegradation. We investigate whether controlled internal micro-variability can insteadprevent representational collapse in a transformer-based dynamical system.Using a Fitzhugh-Nagumo (FHN)-driven transformer as a testbed, we conduct asystematic sweep of internal Gaussian noise σ across three complementary measurementframeworks: (A-1/A-2) hidden geometry metrics (effective rank, collapseindex), (A-3) upper-bound sweep with MSE acceleration analysis, and (B) phaseenergymetrics (Δθ, R(t), phase-state classification).Across all frameworks, we identify a consistent safe edge band at σ = 0.10–0.20 (peak σ ≈ 0.15) in which effective rank is elevated, collapse index is reduced,edge-state proportion is maximized, and prediction error remains bounded. Belowthis band (σ ≈ 0.00), the system is fully collapse-dominated; above it (σ > 0.30),chaos-state proportion increases rapidly while geometric benefit saturates.These findings provide preliminary evidence that internal noise can function asan anti-collapse resource within a measurable band, consistent with the Layer-9Natural Fuzzy Immunity hypothesis in Safe Attractor Architecture (SAA). We donot claim direct proof of edge-of-chaos dynamics; the results are characterized aspreliminary evidence for a safe-edge-like regime.