Deployment Risk Control for Sequential Prediction Systems
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Deployed prediction systems are maintained while they serve: operators recalibrate thresholds, repair inputs, switch to backups, or abstain, often on the system's own feedback. Yet the available controllers certify pieces of this loop — a coverage recursion, a change detector, a single gate — and no unit of accountability spans the interventions, the outputs they serve, and the decisions they drive. We show that deployment risk control is invariance control of one scalar: the running risk-budget balance. Formalizing deployment as a sequential intervention process (SIP), we prove a characterization: a controller is budget-compliant on every trajectory iff it confines this scalar to an invariant interval, iff exponential transforms of the balance are universal-null e-processes; at the boundaries compliance forces the controller's action, so all design freedom lives strictly inside. A constructive two-ledger theorem achieves the invariant: a deterministic ledger keeps the booked risk proxy within of budget on every trajectory, and an intervention ledger bounds interventions fired while their nulls hold, uniformly over stopping times. A certified bridge carries the same ledgers to decision losses, and a gated meta-policy selects among the controllers themselves under one composite anytime-valid budget. We recover ACI, conformal PID, and online CRC as special cases. Under injected degradation on six real streams, the unified policy holds the booked budget at – smaller width than NexCP, the only one of three external baselines that also holds it there; replaying the CDC FluSight and COVID-19 hubs end to end, the certified answer was restraint.
- Primary area:
- learning theory