Engineering Consciousness: A Multi-Circuit Architecture for Functional Synthetic Awareness

Engineering Consciousness: A Multi-Circuit Architecture for Functional Synthetic Awareness (MCAI)

This work presents an engineering blueprint for functional synthetic awareness in AI systems, grounded in network neuroscience. The Multi-Circuit Architecture for AI (MCAI) couples five specialized modules—Sensory, Perceptual, Cognitive (Global Workspace/Planner), Emotional (Value/Salience), and Reflective (Self-Model)—via bidirectional pathways to enable global access, self-monitoring, and context-sensitive integration.

We introduce the Integration Score for AI (IS_AI) as a measurable, optimizable indicator of integrated processing, composed of:

1. KrespK_{\text{resp}}Kresp: algorithmic response complexity (Lempel–Ziv) on binarized activations; a digital analogue of PCI,

2. TEcross\mathrm{TE}_{\text{cross}}TEcross: cross-module Transfer Entropy capturing time-asymmetric, directed information flow with surrogate correction,

3. PglobalP_{\text{global}}Pglobal: graph-theoretic participation coefficient quantifying system-wide integration vs. fragmentation.

The paper provides a falsifiable prediction: targeted perturbation of the Reflective module should reduce metacognition (e.g., confidence calibration) more than primary task performance, with all IS_AI components dropping and recovering in tandem. A two-module proof-of-concept (Sensory–Cognitive) demonstrates that a controlled perturbation jointly lowers KrespK_{\text{resp}}Kresp, TEcross\mathrm{TE}_{\text{cross}}TEcross, and PglobalP_{\text{global}}Pglobal, followed by recovery.

Scope note: we do not claim phenomenal consciousness (qualia). The contribution is an operational framework for engineering and measuring functional correlates of awareness in integrated, self-monitoring AI agents—and for using IS_AI as a safety/control monitor (e.g., detecting low-integration states). The architecture incorporates the core broadcast idea of Global Workspace Theory within the Cognitive module and extends it with an explicit Reflective self-model and a quantitative, optimizable integration metric; it is also compatible with the author’s Multi-Circuit Theory.

Contact: info@jpblankert.com

Keywords

• synthetic awareness

• machine consciousness

• metacognition

• global workspace

• self-model (VAE)

• actor–critic reinforcement learning

• value / salience

• Transfer Entropy (TE)

• Lempel–Ziv complexity (PCI analogue)

• participation coefficient

• directed information flow

• integration score (IS_AI)

• causal emergence

• perturbation testing / ablation

• network neuroscience / NeuroAI

• multi-circuit architecture

• safety metrics for AI

• confidence calibration

• graph-theoretic integration

• reflective module / self-monitoring