Brightness-invariant spectral families converge but clean selection loses recall

Can learned spectral families retain embedded candidates while rejecting likely contaminants?

Four arbitrary-mask Gaussian families from measured flux ratios,187 spatial training sources, covariance floor0.03; score two highest far/mid-IR prototypes before labels. Calibration requires90% precision and80% likely-negative non-selection.

Five analytic controls. Independent native/Gaussian review checks4920 features,74984 source/hidden scalars,410 exports and706 hidden-ratio predictions. Separate scalar review reconstructs thresholds,6050 counts/ratios,498 curves and candidate provenance. Preserve refused all-row fit, corrected mask and derived-variance review amendment.

Converges469updates; hidden-ratio RMSE0.29364dex versus0.65933mean. Test32/90 positive references selected, zero of7 likely negatives selected; one negative is unscorable. Family AUC0.72863 trails brightness0.87179.115 canonical provisional candidates (48Class0/55I/12flat), all existing catalogue references.36bindings,0.81MiB, no downloads.

Next: Use actual common survey bandpasses across field stars, galaxies and protostars; retain missing-normalization cases, compatible dust evidence and substantial independent contaminant support. Freeze before fresh unfiltered survey PR/review-burden evaluation and temporal refinement.

Consumed target-conditioned development; no broad-survey performance or new discovery.34 targets lack normalization;36 lack score. Family recovery35.56%,14/30Class0 and0/9no-PACS. Same-input comparison has only2negatives; calibration also2.32/32 nominal binomial interval89.11–100%; zero/7 nominal upper95% bound40.96%; independence/sampling unverified. No physical class/dust/birth adopted.