Broader noise-aware spectral embeddings learn structure but do not beat the simple prediction baseline consistently
6 Oct 2026 · embeddings
Can unlabelled short-infrared spectral shape retain native missingness and useful structure without using overall brightness?
Register seven-band magnitude contrasts with explicit six-dimensional missing-band operators and shared reference-error covariance. Train on supported D1-D4 sources with exact native ID modulo5 nonzero; exclude internal calibration, D5 and Orion. Freeze primary8families seed1570, two fixed seed checks and one-family baseline before hidden-band diagnostics. No labels, seed selection, downloads or physical classes.
Seven controls pass. Independent reconstruction matches27538magnitude/error cells and1967inventory rows. Scalar Gaussian/initialization/finalEM/inference replay checks334694quantities within1.482e-12. Direct joint measured-covariance prediction replay checks28485quantities within2.298e-10 across1899correlatedcells/431sources. Native embedding serialization matches104251native/provenance and59080embedding cells exactly. All workers observed terminal0;55scientificbindings accepted.
All1967sources retained;1688have shape embeddings and279remain unavailable singletons.845unlabelled training sources and227internal calibration sources remain disjoint. Primary/sensitivity/single fits converge in358/212/499/176updates. SeparateD5diagnostic83sources/334cells: primaryRMSE0.631447vsone-family0.624734 and training-mean1.321323; primarymean log predictive density−0.060550vsone-family−0.291931. Every source carries native/model provenance, observed-band support, conditional colours/covariance and statistical-family responsibilities without physical labels.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| native sources with shape embeddings | 1688 | — | 1967 |
| unlabelled training sources | 845 | — | — |
| separate diagnostic predictive sources | 83 | — | — |
| primary diagnostic prediction RMSE | 0.6314471403182657 | — | — |
| single-family diagnostic prediction RMSE | 0.6247338630037221 | — | — |
| confirmed physical births | 0 | — | — |
Next: Register actual-band guarded ordinary-stellar and infrared-excess/embedded retrieval with matched raw-colour, brightness and one-family baselines, all-requested recovery and review-cost curves before reference inspection. Obtain representative positive/negative calibration and fresh transfer. Preserve all previous candidates, singleton and far-IR follow-up plus physical ownership/membership/temporal birth-region scope. Fullgoalactive;zero downloads,E121zero,12MiBcap/5GiBreserve. E156/E157external updates queued behind pending separate readback approval.
Primary loses RMSE to one-family in every cohort: calibration0.885342vs0.813474;Orion1.234978vs1.177529. It also loses predictive density in both calibration andOrion. Different local solutions and small sensitivity families5.18/7.23effective sources limit interpretation; covariance floors bind. Held-out cells are correlated, only83D5sources qualify, and near-IR availability/geography/epochs/noise/artifacts/blends/galaxies limit transfer. No candidate retrieval, population precision/recall, physical stellar/dust classes, membership or birth is validated; latent missing colours are predictions, not observations.