When Isotropy Fails: Data-Guided Geometry Selection for Genome JEPAs

We study a training-free selector for collapse-avoidance geometry in joint- embedding predictive architectures (JEPAs), using sliced negentropy and a marginal-preserving permutation null. In synthetic worlds, isotropic-Gaussian regularization moves from first to last among four geometries as latents become clustered. Four genome representations exceed their shuffle baselines. With characteristic-function SIGReg and per-geometry weight tuning, held-out-family trait AUROC is 0.712 for EMA, 0.703 for VICReg, and 0.684 for SIGReg over three seeds. These are exploratory rankings: representations do not identify la- tent Gaussianity, and tuning independence is unresolved. For AgenticLS Track I, the tested selector motivates a proposed bounded model-training assistant, not an implemented or evaluated autonomous or recursively self-improving agent.