Alignment Is a Use Problem. The Field Is Still Treating It as a Design One.

Safety benchmarks and safety indexes measure what labs do before deployment. The argument circulating in alignment research now is that this is the wrong moment to measure. Once a model is in the world, the relevant variable is not how it was trained but who has access to its capabilities and through what interface — a distinction that because constraining post-deployment use does not fit neatly into a capability-evaluation framework.