Training-Inference Kernel Contracts: Bounding Divergence in Post-Training and Deployment

In three linesTheoretical paper proposing kernel contracts to bound divergence between training and inference kernels in post-training. Framework specifying acceptable gaps in finite precision with numerical, statistical, and routing clauses. Derives bounds from logit drift to total-variation distance and applies to RL policy-gradient bias.

Summary generated by Claude — human-verified