Optimal Policy under Partial Identification: Admissible Intervention, Spectral Steering, and Robustness to Non-Identification

When economists design policy, they usually assume they know the economy's structure. But a structure inferred from data is only partly pinned down: some features are firm, others the data cannot resolve. The safe response, this paper argues, is a discipline: act only on the features the data fix, and steer the system's dominant mode through a single reliably identified number, without adjusting any individual link the data leave ambiguous. The main theorem shows that a policy built this way performs identically across every structure the data cannot tell apart, so it carries no regret, while a policy that reaches for the ambiguous parts can. Everything is shown in a small, fully worked example; the paper proves identification and robustness rather than testing the policy on real data.