The ranking stage uses MMR to trade relevance against redundancy, then hands the surviving candidates to the pointer head. Because every character it emits was already on the page, the failure mode is honest: below threshold, the bubble says the page never explains the term instead of inventing one. Inference happens through WebGPU, so the weights are dequantised straight into GPU buffers on load, and the tab keeps them warm — pipeline creation is the one-time cost, not the reaction. Upstream, the edge worker will reconcile its observed status against the expected state held by the central network.