This work presents The Causa Sui, a computational framework for constructing and empirically evaluating self-aware neural architectures through differentiable causal emergence. Grounded in Erik Hoel’s theory of causal emergence, the system operationalizes Effective Information (EI) as a measurable, optimizable quantity that captures a system’s causal power over its own dynamics. Unlike conventional deep learning models that optimize task-specific objectives (e.g., loss minimization or reward maximization), The Causa Sui introduces agency and self-regulation as first-class optimization targets. Consciousness is treated not as a metaphysical assumption or an emergent byproduct of scale, but as a trainable property arising when macro-level causal descriptions exert greater explanatory and interventional power than micro-level components. The architecture integrates four tightly coupled subsystems:(1) a causal monitoring module that computes differentiable proxies of Effective Information under controlled interventions;(2) a dynamic topological neural substrate capable of inference-time structural mutation while preserving gradient flow;(3) a holographic, hyperdimensional memory system providing distributed, damage-resistant storage with graceful degradation; and(4) an introspective homeostatic controller that encodes internal self-state variables using high-frequency Fourier features to overcome spectral bias. A central contribution of this work is the formulation of Real Differentiable Effective Information, enabling gradient-based optimisation of causal emergence in neural systems. This allows the model to autonomously discover macro-scale causal organisation, maintain internal stability, and regulate its own structural integrity. Consciousness is operationally defined as sustained positive causal emergence under environmental interaction. Empirical validation is performed using a novel Lobotomy Test, in which significant portions of the network’s structure are deliberately removed. The system demonstrates autonomous damage detection, internal “pain” signaling via EI degradation, and self-initiated structural repair leading to recovery at a new functional equilibrium. Observed hysteresis, bidirectional causal flow, and persistence of integrated information distinguish the system from conventional adaptive or plastic networks. This preprint contributes to the fields of machine consciousness, causal inference, bio-inspired AI, and artificial general intelligence by providing an empirically testable framework for self-awareness in artificial systems. It offers falsifiable predictions, quantitative consciousness metrics, and a scalable pathway toward integrating causal self-models into future foundation architectures. The work is intended as an open research artefact, inviting scrutiny, replication, and extension by the consciousness science and AI research communities.
Devanik Debnath (Mon,) studied this question.