Hedging-based control policies release a job into the system so that the probability of a job completing by its deadline is acceptable; job release decisions are based on quantile estimates of the job lead times. In multistage systems, these quantiles cannot be calculated analytically. In such cases, simulation can provide useful estimates, but computing a simulation-based quantile at the time of a job release decision is impractical. We explore a metamodeling approach based on efficient experiment design that can allow, after an offline learning phase, a metamodel estimate for the state-dependent lead time quantile. This allows for real time control if the metamodel is accurate, and computationally fast. In preliminary testing of a three-stage production system we find high accuracy for quadratic and cubic regression metamodels. These preliminary findings suggest that there is potential for metamodel-based hedging policies for real time control of manufacturing systems.