Synthetic Self as a Minimal Boundary Condition for Stable Recursive Intelligence

We introduce Synthetic Self, a minimal internal reference structure required for stable long-horizon, model-level learning and recursive self-improvement in artificial systems. Unlike interpretations invoking consciousness or agency, Synthetic Self is defined as a purely functional and engineering construct that enables identity continuity, historical cost accumulation, and stability across representational phase shifts. We present Outcome Consequence Backpropagation (OCB) as a concrete architectural framework implementing Synthetic Self through two core mechanisms: the Sequence Blame Tree (SBT), which provides non-vanishing, hierarchical, per-layer responsibility assignment over long temporal horizons, and the Predictive Capacity Forecaster (PCF), a lightweight online predictor enabling proactive regulation of learning dynamics before catastrophic instability occurs. Empirical evaluations across long-horizon reinforcement learning environments indicate that systems equipped with Synthetic Self components exhibit significantly improved stability, reduced collapse events, and enhanced capacity for sustained recursive adaptation. These results support the view that Artificial General Intelligence should be understood not as a future model, but as a stability regime emerging when time, consequence, and internal reference are formally integrated. Correspondence to: s.miksztal@gmail.com