Abstract
A failed network need not be fully diagnosed before service can be restored. The relevant decision is whether available observations support a feasible repair across all remaining physical failure states. We formulate this question for finite networks with admissible path measurements, overlapping repair-validity regions, intervention costs, and explicit post-repair transitions. Building on established decision-region determination, we characterize robust terminal repair sufficiency, construct indistinguishability counterexamples, and present an exhaustive finite-horizon decision recursion. For a restricted single-failure series path with one irreversible repair, the optimal success probability under a uniform prior is exactly 1/n. An independently checked finite-state implementation evaluated 30 synthetic instances, producing 180 policy records; four deterministic heuristics matched the exact Bayesian-success oracle, while a random comparator was inferior on three instances. Experiments on three historical network topologies generated 720 fixed-route synthetic-failure records and 960 sensitivity records. Baseline restoration counts were 12/40 for Abilene, 7/40 for GEANT 2012, and 6/40 for ARPANET 1972. These conditional outcomes do not establish field reliability or the value of adaptive sensing, because the evaluated fixed-route policies use no informative probes. The analysis identifies the physical feasibility, decision criteria, and experimental controls required to assess when network information improves restoration.