Paper proposes feasibility taxonomy for governing AI at inference time

A new arXiv paper argues that current compute-governance frameworks concentrate on training-stage compute thresholds and treat the finished model as the main regulatory object, overlooking risks that arise after deployment. The authors present a taxonomy that evaluates which governance measures applied at inference time are technically and practically feasible to implement. The goal is to help regulators and developers extend oversight into post-deployment operation rather than relying solely on training-based rules.