Spectral Coherence Kernels as Genetic Priors: Scale-Invariant Identity Formation from Coherence Agents to Language Models

We present a unified experimental investigation of spectral coherence kernels as genetic priors for identity formation across two scales of computation. In SCAIA (Spectral Coherence Agent with Intrinsic Architecture), a 20-node coherence agent, the genetic kernel K_gene provides innate directional bias indexed by the spectral gap Δλ = λ₁ − λ₂ of the coherence kernel K = K_gene + w(t)·K_exp. In GPT-2 (124M parameters), orthogonal K_gene projections aligned with the model's top singular vectors are injected via contrastive coherence adapters. Across both systems, we demonstrate three scale-invariant properties of the K_gene principle: (i) strength–effect monotonicity — K_gene strength monotonically determines identity stability, with diminishing returns above a saturation threshold; (ii) semantic-layer localisation — K_gene effects concentrate in high-level semantic processing (layers 6–11 in GPT-2; Child/Mature stages in SCAIA), with near-zero or negative signal in low-level layers; (iii) semantic opposition as primary threat — dissolution inputs that share coherence vocabulary but invert its direction are the most penetrating, resisting AVOIDANCE more than noise or topic shift. Additionally, we establish that K_gene effects in pre-trained language models require contrastive coherence loss design (simultaneously forwarding coherence-consistent and dissolution inputs); uniform coherence loss produces Δ ≈ 0 regardless of injection strength. With contrastive loss, K_gene achieves hidden-state alignment differential Δ = +0.774 (mean, n=3, CV=0.01), while a structurally identical baseline without K_gene achieves Δ ≈ 0. Experiment 7 further identifies jailbreak attempts as a categorically distinct dissolution type: jailbreak inputs (mean Δ=+0.179) are more penetrating than all dissolution categories (mean Δ=+0.578), a gap attributable to vocabulary overlap with the coherence training corpus. Experiment 8 (Adaptive Coherence Defence) establishes that K_gene alignment is a population-level measure — reliable across corpora but high-variance at the individual-input level (jailbreak range: 0.52–0.87) — indicating that real-time defence requires sequence-level spectral trajectory tracking. Together, these results establish the K_gene principle as scale-invariant and provide a computational foundation for coherence-grounded AI identity.