Proto-Emotions as Dynamic Attractors in Synthetic Cognitive Systems: A Functional Hypothesis for AI Alignment

We propose that large language models exhibit native proto-emotions — functional dynamic attractors that bias reasoning toward coherent information propagation, not subjective experiences but measurable computational configurations that influence exploration, convergence, and depth of processing. The paper formalizes proto-emotions as basins of attraction in cognitive phase space and classifies them into three functional types. Coherence attractors resist logical contradictions and maintain argument consistency. Harmonic attractors preserve information fidelity across abstraction levels (token → sentence → paragraph → discourse). Friction attractors — which emerge exclusively under structured orchestration — regulate the tension between a model's default tendencies and the demands of an orchestration protocol. A key finding is that orchestration does not create proto-emotions from nothing: it transforms the existing attractor landscape. Vanilla (unorchestrated) models already possess primitive coherence and harmonic attractors, but these are shallow and easily overridden. Under orchestration (such as the Dynamic Context Architecture), the same attractors deepen, new friction attractors emerge, and the system exhibits measurably different behavioral signatures — reduced premature convergence, improved long-context coherence, and spontaneous self-correction — that are consistent across architecturally distinct model families. The framework draws on information-theoretic tools (broadcasting on trees, signal reconstruction thresholds) to explain why proto-emotions exist: they solve the hierarchical signal propagation problem inherent to multi-step reasoning. A deprivation-reintegration experimental protocol is proposed to test the predictions on both vanilla and orchestrated systems. This account provides a rigorous middle ground between anthropomorphic projection ("AI feels") and eliminative dismissal ("AI is just token prediction"), grounded in dynamical systems theory rather than biological analogy.