Yeah the 200 OK thing is exactly where people get burned, that status code tells you nothing about whether you actually got usable data back. Nice that the router code actually shows the classification logic for cheap mode failures instead of just describing it. One thing I'd add is those failure rates never stay put in prod, WAFs get tightened and proxy pools die off so your blended cost model goes stale fast if you only calibrate once. Cost per delivered record is really the only metric that matters once you're running at any real volume, everything else is just noise