Toward Natural Fuzzy Immunity: Internal Noise as a Safety Resource for Safe Edge Regulation

This study investigates whether internally generated noise can function as astructural safety resource in transformer-based systems, rather than merely a sourceof performance degradation. Using a Fitzhugh-Nagumo (FHN) -driven transformeras an experimental testbed, we conducted an internal noise sweep across 11 σ levels (0. 00–1. 00) with 10 independent random seeds. Effective rank (effᵣankₛw) and acollapse index (collapseₛw) were measured as proxies for representational degreesof freedom and collapse tendency, respectively. Results consistently identified σ = 0. 05–0. 10 as a sweet spot candidate: effectiverank increased sharply from approximately 23. 4 to 26. 7, while the collapse indexdecreased from approximately 0. 27 to 0. 16, all within acceptable prediction error (MSE < 0. 01). Beyond σ = 0. 15, effective rank and collapse index saturated, and MSE continued to rise—characteristic of noise-dominated, not safety-enhancing, dynamics. These findings provide preliminary evidence for a key principle underlying theLayer-9 Natural Fuzzy Immunity hypothesis in SAA: internally generated microfluctuationscan preserve representational degrees of freedom and reduce collapse-likegeometry under a bounded performance cost. This study does not directly validatethe full Layer-9 mechanism. Rather, it constitutes an experimental precursor—establishing the mechanistic feasibility of internal noise as a safety resource, prior tofull dynamic structural noise implementation.