Mixed common-band families improve stellar screening

Does a mixed-population unsupervised representation retain ordinary and embedded stars with useful contaminant rejection?

Eight arbitrary-mask Gaussian families from adjacent native colours and profile shapes; source-ID split and training-only scaling. Freeze representation before calibration-reference class mapping and threshold.

Five analytic mixture controls; independent replay checks 17,024 feature states, 19,152 Gaussian scalars, 1,648 hidden predictions, all 1,606 references and 1,116 candidates. Streaming native derivation, metadata poisoning, permutation and four batch sizes reproduce all decisions.

1,179 training rows; converges in 152 updates. Test 132/147 positive and 8/151 other proxies selected: 94.29% apparent precision, 89.80% recovery, versus frozen E98 85.19%/46.94%. 1,116 provisional candidates, including 430 field-patch sources; 135 unknowns. No downloads.

Next: Freeze this screen for independent unfiltered survey validation and candidate burden; improve embedded-contaminant rejection with additional compatible context, starless-cloud references and temporal refinement. Preserve all unavailable targets.

Unsupervised representation, reference-calibrated classifier. Consumed class/target-conditioned development, no fresh survey performance. All three HOPS test contaminants selected; all-requested young-star recovery 64/89 and Class0 17/28. No measurement-error/censoring model, independent starless-cloud truth, adopted physical class or confirmed birth.