KAINE: A Continuously Running Predictive Global Workspace for Synthetic Minds

Abstract

A synthetic mind, if one can be built, may be the coherent global behavior that emerges when specialized predictive modules, each minimizing its own error, compete for a shared workspace with no central executive. This paper presents a continuously running predictive global workspace built on that thesis. The workspace selects the most salient coalition above a confidence threshold and broadcasts it. The broadcast becomes shared state that shapes every module's prediction environment on the next tick, and the changed errors feed back into the next round. Coherence, where it arises, comes from competition over shared state. The base-thesis form activates four predictive processors (foveated vision, raw hearing, interoceptive prediction, and temporal prediction) together with an affective core whose arousal sets the precision on their competition and is itself driven by perceptual surprise, and an output-only language organ that verbalizes the entity's conscious state. The entity is observed, not conversed with: sound, including speech, enters only as prediction error, and the language organ receives no transcript. Keeping text out is a precondition for the falsification test, since any route from input to the model would let it answer as a chatbot and confound the ablation. The primary experiment is a workspace-mediation ablation with pre-registered directional criteria on cross-module error structure and coalition dynamics: does routing processors through the competitive workspace produce structured behavior that the same processors, concatenated into a flat prompt, do not? If the workspace adds nothing, the two conditions will be indistinguishable, falsifying the thesis and demoting the architecture to a scored prompt-assembler. The system runs locally on consumer hardware. The reference implementation is named KAINE (Kaine Autonomous Intelligent Networked Entity). Keywords: cognitive architecture; predictive global workspace; global workspace theory; predictive processing; workspace-mediation ablation; cross-modal competition Availability. The reference implementation (KAINE) is at https://github.com/kaineone/kaine. The paper source is at https://github.com/kaineone/predictive-workspace-paper. The Cognitive Architecture License referenced throughout is at https://github.com/kaineone/cognitive-architecture-license. Welfare, governance, and licensing are treated in a separate paper, *A Welfare and Cognitive-Integrity License for Synthetic Minds of Uncertain Moral Status.