From Prediction to Self: Developmental Conditions for Agency in Minimal Neural Systems

How does a system that merely predicts the world come to distinguish its own causal influence from everything else? We trace this transition in a minimal 192-dimensional GRU through 40 controlled experiments arranged as a developmental sequence. Starting from a system with no action and no self-representation, we add components one at a time — causal action loops, proprioceptive channels, learnable action policies — and track, at each stage, whether the system can distinguish self-caused from world-caused changes in its observations.

The developmental path reveals four necessary conditions that must be satisfied in strict order: (1) persistent state that forms stable attractors, (2) a causal action loop linking the system's output to its input, (3) proprioceptive feedback that makes implicit causal knowledge explicit, and (4) asynchronous awakening — perceptual learning must consolidate before action learning begins. We propose agency gain (A = Err_world - Err_self), the predictive advantage of knowing one's own action, as a metric to track this developmental process. In the final configuration, the self-aware predictor consistently outperforms the self-blind predictor across both periodic (sinusoidal) and chaotic (Lorenz) environments, and the metric survives ablation of all auxiliary components. Only forward-sampled action selection produces meaningful agency gain; two gradient-based alternatives degenerate.

Equally significant are the 12 falsified hypotheses that map where development stalls: predictive coding alone does not produce self-representation, passive memory cannot sustain post-action state, complex probes cannot extract what is not encoded, and awareness and intention cannot be co-learned. These negative results delineate the boundary between systems that predict and systems that know they are the ones predicting. Moreover, the system sustains self-representation only when it is causally useful: after the external training signal is removed, the causal agent retains its encoding (94.9%) while a statistically-matched control collapses to chance (53.9%).