The Thinker-Feeler Architecture: A Model-Agnostic Affective Feedback Layer for Emotionally Grounded AI

Why does AI feel hollow? Not because it lacks intelligence, and not because it lacks data. Because it is architecturally incapable of feeling anything before it speaks. Contemporary Large Language Models attempt to simulate both rational analysis and emotional resonance within a single monolithic neural network. This paper argues this is the fundamental reason AI outputs feel mechanically correct yet emotionally empty — a phenomenon termed the Aura Gap. The Aura Gap is not a data problem or a scale problem. It is an architectural problem rooted in a single design decision: current AI systems apply emotional tone to an answer after it is generated. They do not pass through any internal affective state that constrains the answer before generation. This paper proposes the Thinker-Feeler Architecture: a decoupled dual-system design that separates a language-based analytical engine (System 1, the Thinker) from a non-linguistic, model-agnostic, finite-state affective simulator (System 2, the Feeler). The two systems communicate through a Virtual Endocrinology Layer — a hormone-vector protocol inspired by human neurochemistry — before any output is produced. The paper formalises the interface specification, presents an implementation sketch and ablation proposal, identifies three core open problems (Binding, Grounding, Bootstrap), and carefully delimits the scope of claims. The architecture is speculative but implementable. It does not claim to produce consciousness or genuine subjective feeling. It proposes a structural path toward AI systems whose outputs are conditioned by internal affective state dynamics rather than by statistical patterns of what emotional language looks like.