JudgeGPT launches local LLM-judge benchmarking with GPU telemetry

This is a useful step toward more auditable local eval workflows, especially for teams that distrust one-shot judge prompts.

• –Behavioral scoring anchors across five criteria help reduce small-model leniency drift compared with naive judge setups.
• –Runtime-configurable judge model and system prompt make judge-disagreement studies practical without editing configs.
• –Self-family bias warnings and optional human score blending add guardrails against overtrusting automated scores.
• –Real-time Metal/ROCm/CUDA telemetry plus SQLite history make performance-quality tradeoff tracking much easier over time.