The minimax-m2.5 case is the one that would bite you in prod - it accepts every switch, returns 200, and just ignores all of them, so nothing tells you it did nothing. We got burned by something close: a router model silently choosing different temperature/top_p depending on which backend it picked that request, and we only caught it because a batch of outputs came back oddly uniform. Three-sample runs would've missed that too, same as your minimax retraction. Did you check whether the router: targets report reasoning_tokens more honestly because they're just passing through the real model's usage object, or was that just how your sample landed?