Admissible Portfolio Optimization: Information Eligibility Before Portfolio Choice

Portfolio optimization normally treats the conditioning information as fixed and optimizes only the allocation. This paper studies the prior problem of which information representations are eligible to define the portfolio problem. A decision contract induces an eligible information domain, an admissibility order selects within that domain without using portfolio performance, and conditional portfolio optimization is performed only afterwards. The first main result shows that decision value and information eligibility are distinct: a finer representation can improve optimal decision value while violating the contract, so performance alone cannot certify eligibility. The second shows that joint optimization over information and portfolios can generate a pointwise efficient envelope that is not the frontier of any single information representation; representation consistency holds exactly when one representation minimizes the envelope throughout the target range. Conditional Markowitz choice is recovered when the eligible information domain collapses to one equivalence class, and classical Markowitz when that representation is trivial. Controlled experiments exhibit strict performance preference for ineligible look-ahead information. In a frozen market experiment with 127 candidate drivers, the performance-selected representation fails the independently predeclared statistical eligibility gate in all 81 rolling windows. The gate is deliberately stringent: APO finds an eligible representation in only 6 windows, and no tolerance is changed after observing the results. APO does not imply that eligible representations maximize portfolio performance; it separates the legitimacy of the information domain from optimization within it.