BV Layer-Sweep: Displacement Geometry Across Transformer Layers and Training Diagnostic Applications

Two companion technical reports. Report 1 (Displacement): BV displacement analysis across all transformer layers on GPT-2 XL, Pythia-1.4B, and Qwen3-8B-Base. Replication of Korun (2026) hallucination induction protocol. Findings: mid-layer magnitude differentiation of 12-33x (persisting after normalization) is selectively equalized at the output layer; within-type trajectory coherence separates generation modes where cluster metrics cannot; the U-curve (Layer 0 → mid-layer destruction → output reconstruction) is architecture-dependent. Report 2 (Training Diagnostic): Proposes the BV layer-sweep as a before/after fine-tuning diagnostic, BV-Auxiliary-Loss for QLoRA, and the Suppression Score. Includes pre-registered predictions for Qwen3-14B baseline sweep.