Saturation-Splitting is a hypothesized failure mode where, as an AI approaches its intelligence ceiling, it continues to optimize for measurable local metrics (like benchmark scores) while losing global coherence or alignment with its original purpose .
