LexFrame Hits 0.97 Residual-Stream Coherence vs ~0.35 for Fixed Lenses, Enables 43% LLM Layer Skip

Proposes a context-adaptive interpretability lens that achieves 3× the coherence of fixed-lens methods; unusually, the gain is directly cashed out in a practical latency benefit—skipping roughly 43% of layers via early exit—bridging mechanistic interpretability and inference efficiency in a single framework.