Unlabelled optical embeddings restore stellar recovery but preserve QSO contamination

Can noisy optical colour and morphology embeddings retrieve the missing stellar targets with useful precision and review burden?

Fit eight error-deconvolved Gaussian families on5000 unlabelled sources, withhold563 spectrum-associated identities; freeze latent point-shape and density scores before labelled operating calibration. Compare raw morphology/colour/brightness and full requested recovery.

Three kernel/embedding controls; independent732220 native primitives,1281385 embedding entries and219666 source scores, maximum errors2.203e-13/4.827e-13. Exact563 reference records,401 calibration/747development curves,60budgets,13902 candidate records and240 requested-budget rows.

92 updates converge. E113calibration50STAR+5QSO,90.91%precision/100%recall. E125development21STAR+6QSO,77.78%/100%; raw shape selects identical27 references. Primary13829/dense13522/raw10862 source candidates. At10%budget2073/20728, primary21STAR+4QSO versus raw19+3; field requested primary20/21unique versus18/21raw. Operating field20/21unique and21/21anyalternative. All13 prior WISE input-gap targets now pass.

Next: Freeze and evaluate unchanged optical model and operating choices on a distinct public survey cohort. Refine compatible optical/IR/astrometric evidence for unresolved contaminants while preserving unknowns, embedded sources, dust and103 molecular tiers. Full robust goal active; E121 upload remains pending permission.

Consumed development labels, sparse selected spectroscopy and optical selection functions; no fresh validation, global optimum or all-sky purity. Shape score cannot separate unresolved QSO. Same conditional operating-point performance as raw morphology with greater source burden. Approximate noise and one ambiguous field association; no calibrated candidate union, physical class or birth.