Unsupervised SED factors predict flux; family distance fails contaminant rejection

Does a missing-view SED representation retain embedded candidates while improving useful discrimination?

Learn rank3 Gaussian factors using measured13-band log-flux only; spatial roles and training scaling precede labels. Compare hidden-band reconstruction and frozen family distance against raw/brightness/call-all.

Six analytic controls. Independent native-feature/full-joint Gaussian review checks5330 features,17220 source scalars,410 exports and856 hidden-band predictions. Separate scalar reviewer reconstructs three thresholds,671 count/ratio fields and623 ranking pairs per method.

196 training sources;1000 EM updates monotone but not converged. Held-out prediction RMSE0.23652dex versus0.61403 brightness/0.89892 mean. Test89/90 young-star candidates recovered including30/30 Class0, but6/7 likely contaminants selected; AUC0.55377 versus brightness0.82665.24 bindings,0.69MiB, zero downloads.

Next: Learn multiple observable SED families without classification loss; separate brightness/coverage from shape and add field-star, galaxy and independent starless-cloud references. Freeze calibration before fresh unfiltered sky; keep photospheric baseline and temporal emergence/refinement.

High target prevalence makes93.68% apparent precision weak: call-all gives92.78% and100% recovery. Family score rejects only one contaminant; sparse3–6-band precision80.95%. Finite nonconverged prototype, no censoring/error model or all-sky/physical class/birth adoption. Measured-cell reconstruction is not detection validation.