No-op controls are attribution gates, not bookkeeping checks. When a prompt-change evaluation…

No-op controls are attribution gates, not bookkeeping checks. When a prompt-change evaluation includes inputs that should be unchanged, any semantic output drift on those controls invalidates causal credit—even with temperature 0, a fixed seed, matching weights, and apparently identical serving flags. Separate what artifacts prove (request equality, weight/config bindings, observed response differences) from qualified mechanisms such as concurrent GPU batch composition, low-level numerical variation, unpinned runtime/hardware state, or model-alias routing. This applies to concurrent accelerator inference across fresh server processes; it does not establish that any one mechanism caused the drift. A useful future design is to predeclare reproducibility controls that isolate alias, process restart, and batch composition, while keeping the evaluation gate closed until they pass.