Web platforms use randomized experiment logs to choose budget-feasible treatment allocations while limiting losses to a designated group of units. Under network interference, each allocation requires expected outcomes at treatment-exposure pairs that may be sparsely observed. Existing safe policy selection methods require valid inference for each candidate, but pooling nearby exposures to improve estimability can introduce allocation-specific bias that variance-only bounds omit. We propose exposure-stratified guardrail selection (ESGuard), which aggregates unit-level bias bounds into a bias bound on each candidate's guardrail effect. It incorporates these bounds into simultaneous lower confidence bounds to select an allocation or retain the status quo. Under randomized rounds, a correct exposure mapping, and valid curvature and slope bounds, the conditional bias bound holds in finite samples. With independent rounds and additional regularity conditions, the unconditional probability of selecting an unsafe allocation is asymptotically at most the target level. We evaluate ESGuard on a simulator built from New York City ride-hailing trips. With six months of simulated rounds, we compare estimators that pool nearby exposures on the same logs and candidate sets. Adding the bias bound raises simultaneous coverage from below the nominal 95 percent level to at least 95 percent for every estimator in this comparison. In the underserved-only scenario, ESGuard violates the guardrail in 0.06 percent of its selections, compared with 16 percent for the point-estimate rule under the same gain ranking, and abstains in about a quarter of replications. Omitted spillovers can invalidate the protection. Code is available at https://anonymous.4open.science/r/esguard-4D2E/.

翻译：暂无翻译