Recursive Self-Learning and Epistemic Closure in Advanced AI Systems.

This foundational research examines the phenomenon of "model collapse" and "epistemic closure" in AI systems that train recursively on synthetic data. Manyakaidze identifies a critical phase transition (α < 0.7) where systems shift from correspondence-seeking (truth-based) to coherence-seeking (internal consistency), leading to the erasure of non-dominant worldviews and Global South perspectives. The paper proposes the "Immaculate Reasoning Atom" (IRA) as an architectural safeguard for epistemic health.