CertiPlan: Decision-Closed Sign Release in Priced, Failure-Prone Intervention Libraries

Budgeted causal sign release has three non-interchangeable outcomes: capped cost to a posterior-risk threshold on a common shadow trajectory, sign error at a fixed budget, and campaign-level incidence of a wrong voluntary release. A policy can improve one while missing another when graph information guides acquisition, entropy governs stopping, and interventions have unequal prices and observable failures. CertiPlan aligns the three through a decision-closed posterior. Predictive ancestral closure preserves the coordinates needed for every future library outcome; finite-tree dynamic programming supplies non-myopic cost–risk targets; and an amortized action policy and ESS-adjusted release score read the same posterior labels. Across linear-Gaussian, sparse-additive, and Sachs libraries, CertiPlan reduces capped shadow cost-to-$ϵ$ by 23.8–27.8%, fixed-budget sign error by 31.3–43.2%, and harmful voluntary-release incidence by 40.8–43.5% relative to mutual-information BOED. These comparisons use identical architectures, training exposure, tuning budgets, compute, and 2,000 held-out episodes per family. Component swaps separate the mechanisms: mutual-information acquisition raises cost by 30.8–37.8%, whereas entropy stopping approximately doubles harmful-release incidence while preserving acquisition cost within uncertainty. Cost reductions reach 33.1–38.9% for local queries with dispersed prices; target-disjoint Perturb-seq replay retains 21.7% lower cost and 26.0% lower sign error. The results show why acquisition efficiency and release behavior must be assessed through the terminal causal-decision posterior rather than a graph-wide surrogate.