Anytime-Valid Risk Pricing for Runtime Compression in Serving Stacks

A new arXiv paper (2608.15810) addresses the lack of soundness guarantees in runtime compression of serving state, where systems trade quality for capacity without priced risk. The authors demonstrate that traditional union-bound approaches exhaust their risk budget on every long request in a production serving stack, affecting 100% of requests. They propose an anytime-valid, physically accounted ledger that maintains bounds at every admission call, validated on 352,333 live traffic calls. In a pre-registered confirmatory round, this method halves the exact-fallback rate at matched risk (0.30 to 0.14). The paper also introduces a machine-checked design law (TV <= tanh(a_q w_thr)) that converts served-TV targets into threshold knobs, along with a three-layer audit of its implementation, including an operator-norm query envelope. This work provides a principled framework for pricing risk in compressed serving systems, offering a more reliable alternative to existing methods.