The Spontaneous Legibility Assumption: Output-Loop Attractors and the Illusion of AI Self-Modeling > ER1: The Chi Recovery Experiment

Empirical Record 1 of the Three Primitives Framework (3primitives.io). This record tests the Spontaneous Legibility Assumption: the widespread presupposition that transformer architectures will naturally self-organize toward internal legibility, stable self-modeling, or coherent self-reference under load. Using a controlled recursive pipeline on GPT-2 Medium (Protocol v3.0, recursive depth ρ=1-50), we demonstrate that coherence recovery occurs exclusively at structural output boundaries. Semantic self-reference without structural forcing produces monotonic decline. Four competing explanations were pre-registered; three were eliminated. The empirical baseline for spontaneous legibility is zero. Red-teamed by Claude (Anthropic), DeepSeek, Gemini (Google), and Perplexity.Full corpus: https://3primitives.io/formal_records/