Calibrated LLM-as-judge: how I made my LLM give honest 4/10 scores instead of always-an-8