The decision is made per resource, and it is made on measurements taken from inside China rather than on a rule of thumb. That matters because the naive approach — route everything through a China delivery layer — makes some things worse. Measured fact: sampling every ten minutes for twelve hours from a probe inside mainland China, two of five common third-party resources were faster requested directly than through an additional hop, by roughly 170 to 200 milliseconds each. Those should be left alone, and a layer that rewrites every reference by default pays that penalty on all of them. Meanwhile the resources that genuinely need changing fail in a way that speed tuning cannot reach: across 280 enterprise sites measured from inside mainland China, 34% finished none of three attempts and roughly two thirds of those had already answered and painted before stalling. So the useful question about any given request is not «is it slow» but «does it complete, from where our visitors actually are» — and answering it requires having measured both paths. The delivery arrangement that removes the border crossing is also what brings the ICP filing with it, which is why the two are planned together rather than in sequence.