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.
Rodney Manyakaidze (Sat,) studied this question.