How do you adjudicate between research wanting to scale...

If capacity must be held fixed as the corpus grows, the model size (P) must scale with N. Since decode steps stream P, latency becomes linear in corpus size. At 11B parameters, the latency for a multi-token identifier is significantly higher than a standard ANN scan. Unless identifier assignment improves enough to allow smaller models to compete, the generative approach may be better suited for candidate generation rather than as a total replacement for traditional retrieval.