Geometric Prediction of Economic Behavior II: Held-Out and Cross-Domain Validation

A companion paper introduced a geometric decision model in which economic choices are displacements on a nine-dimensional manifold evaluated by a Mahalanobis cost metric~ , and reported an in-sample benchmark of sixteen aggregate targets. The natural question that benchmark raised---whether the shared metric is a flexible representation or a predictive structure---is answered here with held-out, individual-level, and pre-registered tests, and with a designed falsification instrument. plus a domain-specific magnitude raises transfer to 97--100\ Third, four claims that is low-rank, hashed before data, survive on three held-out domains---social-preference games, a 19-country prospect-theory replication, and a six-language panel---where a rank-1/rank-2 precision matrix matches the full metric, beats the diagonal restriction, and transfers at 99\ Fourth, a design constructs matched problems identical under every scalar projection but differing on one non-monetary coordinate, forcing the entire scalar class to predict no difference; run pre-registered on a panel of six language models (216 persona-conditioned agents, 83,682 forced choices), five of seven contrasts move in the predicted direction in all six models, and the two pre-registered failures are reported. Fifth, both failures localize to the cost-dependent temperature, which a fixed information price motivated by rational inattention repairs, improving held-out fit. Scope is explicit: the projection-gap evidence is from language-model populations, not humans; the low-rank result is gain-domain; and the confirmatory human study these results power is future work, not a claim. Target venue: IEEE TCSS (under review, TCSS-2026-03-0366). Author preprint deposited for archival and citation. Draft — pending author review.