Risk-Controlled Release and Audit of Machine-Generated Galaxy Morphology Catalogues

Amogh Shrivastava

Abstract: Galaxy morphology catalogues for the Vera C. Rubin Observatory?s Legacy Survey of Space and Time will require machine-generated labels at scale, but a model score alone cannot justify releasing a label. Existing studies have already combined machine learning, citizen science, active learning, uncertainty estimation, and human review. This paper addresses the downstream catalogue decision: when may a catalogue release a machine label, when should reviewers inspect it, and when should the catalogue abstain? We define a classifier-independent policy based on predeclared observational strata, random audits, one-sided exact binomial bounds, post-release monitoring, catalogue rollback, and label provenance. The policy converts a stated unsafe-label ceiling into an explicit audit requirement. With no unsafe outcomes in a fixed certification sample for one stratum, the 95% one-sided Clopper-Pearson rule requires 59, 149, or 299 reviewed entries to support ceilings of 5%, 2%, or 1%. We train no classifier and analyse no survey data. This analysis supplies an analytical catalogue-release policy for later testing with Rubin or precursor-survey data.

Keywords: astronomical catalogues; citizen science; galaxy morphology; human-in-the-loop learning; machine learning; selective classification; Vera C. Rubin Observatory