Decision-Aware AI Must Replace Static Classifiers to Unlock LSST Science

The Vera C. Rubin Observatory's LSST will issue roughly 10 million nightly alerts, a volume that exposes the limits of current transient pipelines built around one-shot classification. A June 2026 arXiv preprint by Bom and colleagues reframes the problem as sequential decision-making under uncertainty rather than static labeling, advocating foundation models for evolving source representations paired with decision-theoretic policies for follow-up allocation. This preprint remains conceptual, offering no empirical benchmarks, training datasets, or sample sizes, and therefore carries the inherent limitation that performance claims cannot yet be quantified. Related work on the Zwicky Transient Facility alert stream (Graham et al. 2023, AJ) already showed that pure detection systems overload human follow-up queues by factors of 5-10 within two years of operations, a pattern LSST will amplify by two orders of magnitude. A second synthesis with the LSST overview (Ivezić et al. 2019, ApJ) reveals an under-appreciated bottleneck: the survey's cadence and depth will generate partially observed light curves whose scientific value depends on real-time choices about which objects receive spectroscopy or multi-band photometry. The Bom framework correctly identifies that embedding these choices inside human-supervised agentic loops can redistribute scientific agency, yet it underplays the governance challenge of encoding utility functions that avoid concentrating discovery power among institutions with privileged access to follow-up resources. Without auditable value metrics tested on realistic survey simulations, the proposed shift risks optimizing for easily measured quantities rather than rare but high-impact events such as kilonovae or interstellar objects.