I wonder if the degradation of gpt2.5 is because it doesn’t randomize the seed with every edit. Drift gets dramatically worse if you keep the same seed (seen on local models like qwen and Klein edits )