Pick, Don't Blend: Label-Free Selection of Models, Layers, and Seeds for Financial Asset Representations

Financial temporal systems increasingly offer many candidate representations---architectures, pretrained time-series models, layers, and random seeds---but selecting among them with a downstream validation target can consume scarce labels and invite temporal leakage. We propose a label-free alternative: score each asset representation by its calibrated agreement with a leave-one-out consensus of the other candidates, then select the maximum- or minimum-agreement candidate according to whether the task requires shared or residual cross-sectional structure. Across five equity panels, 16 representation families (including three frozen time-series foundation models), 38 candidates per problem, and nine tasks, the label-free ranking agrees with the downstream ranking at Spearman $\rho=0.93$. Selection recovers 48--100\% of the random-to-oracle gap on every task, identifies null embeddings, and selects network layers in 40/40 autoencoder and 38/40 transformer trials. We also report sharp boundaries: heterogeneous or very small panels, mixed-structure tasks, and pool composition can defeat the rule. The result is a leakage-aware selection primitive for reliable temporal representation systems.