Taxonomy-Shaped Mixture-of-Experts Routing Atypicality for Zero-Shot Out-of-Taxonomy Transient Discovery in Rubin/LSST and ZTF

Real-time alert streams from the Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) and the Zwicky Transient Facility (ZTF) will deliver millions of multiband photometric light curves per night. While sparse Mixture-of-Experts (MoE) foundation models such as Astro-MoE achieve strong closed-set classification across known transient families, standard out-of-distribution (OOD) detectors—including Maximum Softmax Probability (MSP), Output Predictive Entropy, and penultimate-layer Mahalanobis distance—fail catastrophically when unseen physical transients (e.g., Kilonovae, Intermediate Luminosity Optical Transients, Pair-Instability Supernovae, and Fast Blue Optical Transients) project into the convex hull of known classes after sequence pooling. We introduce Taxonomy-Shaped MoE Routing Control and a zero-extra-training anomaly statistic, MoE Routing Atypicality (A_route). By supervising internal router trajectories with stabilized Light-Curve-Level Taxonomy Mutual Information (J_tax) and smooth-flux-gradient Temporal Consistency (L_temp), we increase the Within-vs.-Cross Taxonomy Cosine Similarity Gap from 0.000 to 0.528. Across held-out rare transient classes on the LSST 6-band (u,g,r,i,z,Y) ELAsTiCC/PLAsTiCC and FLARE benchmarks, A_route achieves 98.0% Recall at 1% False Alarm Rate (FAR) (AUROC = 0.9993), outperforming Output Predictive Entropy (1.8% Recall @ 1% FAR, +96.2 pp) and Penultimate-Layer Mahalanobis Distance (9.9% Recall @ 1% FAR, +88.0 pp).