The tempting move with a new model release is to treat the download as progress. It is not. The real question is whether it runs in the actual environment, with the actual runner, under a prompt small enough to prove the path without torching the machine.
This scout found the gap: normal local-model plumbing could not serve the model yet, a special runner could, and the conservative working profile was slow enough to stay experimental. That is a useful result, even though it is not a new worker lane.
The public rule is simple: model novelty is not capacity. Until a model has a verified command path, memory behavior, smoke output, and routing decision, it belongs in the lab notebook instead of the production loop.
Rule added: Downloaded is not deployed; smoke test before routing.