Deployment-Matched Allocation of PBE, mBJ, and HSE06 for Band-Gap Screening Under Chemical and Protocol Shifts

Band-gap screening requires a calculation policy because agreement and cost vary with chemistry and computational protocol. The observation target is one literature gap compiled for each benchmark record across optical and fundamental measurements, temperatures, and sample conditions. A depth-two decision tree using band-edge orbital character reduced mean absolute deviation from this compiled target by 35.5 meV (95% confidence interval, 19.9-52.4 meV) relative to uniform mBJ while assigning HSE06 to 18.7% of 461 compounds. No equal-count random policy matched the router among 50,000 draws (p = 2.0 × 10^-5). Complete retraining under global target shifts up to ±100 meV and under 25-100 meV record- or source-correlated jitter retained a positive allocation gain. Under leave-element-out transfer, the compound-level gain remained 24 meV (95% confidence interval, 10-38 meV) and was concentrated in chalcogenides and oxides. At the same HSE06 count, gap- and composition-based learned rules matched the orbital router. The framework therefore links allocation, admission, observation-target sensitivity, and chemical transfer in one deployable screening rule.