Judge drift is one of the most under-discussed failure modes in eval, so it is good to see it named directly. The trap I hit is that when the score drops you cannot tell if the system got worse or the judge did, unless you pin the judge model version and keep a small fixed set you re-score every release. I wrote a related piece on why I stopped scoring only the final answer for agent apps: kartiknvjk.hashnode.dev/evaluating-openai-agents-… . How are you separating judge drift from genuine model regression in production?