A Predictive Global Neuronal Workspace for a Continuously Running Synthetic Mind: Architecture and a Falsifiability-First Evaluation Framework

Today's AI with memory keeps a language model at the center: memory is just retrieval, feeling is a prompt, and between turns nothing continues. Classical cognitive architectures take structure seriously but predate modern learned models.

Treat a mind as an ongoing process, not a model. Sixteen modules run continuously and compete through a shared predictive workspace; the most surprising, confident estimate is broadcast to the rest, and the language model is only its voice.

It runs the brain's predictive-workspace theory as a live system with no central controller, where one loop both remembers and corrects itself and safety lives in the architecture, not the model's weights. Every claim is tied to a falsifiable test.