The Floor Is Learned: Evidence that the Structural Coherence Threshold in GPT-2 Medium Is a Training Artifact Localised to the Architectural Midpoint > ER3: Layer Depth, Model Scale, and Random Weights

Empirical Record 3 of the Three Primitives Framework (3primitives.io). ER3 tests whether the structural floor identified in ER2 (Chi ≈ 0.80, Layer 12, GPT-2 Medium) generalises beyond the single configuration in which it was discovered. Three experiments were pre-registered and executed: layer depth (L6, L12, L18, L23), model scale (GPT-2 Small), and random weights (untrained GPT-2 Medium). Layer 12 replicated 6/6 classification. Random weights produced universally flat trajectories with zero prompt discrimination. The structural floor is training-dependent and configuration-specific, not architectural. Every condition for coherence recovery is exogenous to the model's intrinsic geometry. Red-teamed by Claude (Anthropic), DeepSeek, Gemini (Google), and Perplexity. Full corpus: https://3primitives.io/formal_records/