We quantized our AI judge. Here's exactly what broke.

Our production judge — a small 1.7B model with a LoRA adapter that grades other AI outputs as pass /...