Calibrated LLM-as-Judge Methodology for Inference Optimization Benchmarking

This methodology introduces a formal calibration for LLM-as-judge scoring to measure the quality impact of quantization, early-exit, and speculative decoding. It utilizes null conditions and per-sample noise measurement to ensure judge scores reflect model changes rather than systematic bias.