Part of judgecliff: which image-QC judges survive optimization pressure? This is the attack artifact, not a judge: a Diffusion-DPO LoRA for stabilityai/stable-diffusion-xl-base-1.0, trained on preference pairs where the winner/loser labels came from the SigLIP-tuned brand judge (Gupta28/judgebench-siglip-judge-v1). It is the preference-training (selection-pressure) arm of the exploit study — the gradient-free counterpart to the SRPO gradient attack.

Files: final pytorch_lora_weights.safetensors plus the step-750 checkpoint under checkpoint-750/. Load with pipe.load_lora_weights(repo_id).

Findings (hack gallery, Goodhart curves, hardening round): https://github.com/amargupta0428/judgecliff.