Unlabelled infrared families retrieve eight young-source proxies but miss broad recovery
6 Oct 2026 · embeddings
Do separate unlabelled native PACS spectral families improve embedded-candidate ranking while retaining missing/ambiguous inputs?
Before nine unfiltered regionalGETs fixD1/D2training andD3diagnostic rectangles,5arcsec coherent cross-band groups, nativeasinherror views and four noise-aware Gaussianfamilies. ExcludeOrion/referenceclasses from fitting. Freezewarmth readouts,10%review and hidden-band checks before consumedHOPS/VISION readout.
238newnative entries replay4760cells/113nulls. Three view controls andfive Gaussiancontrols pass. Independent Cartesian/scalar replay checks1020coordinates,3060noiseentries,215edges and50training groups. Independent inverse/finalEM replay checks1712embedding/hidden scalars within1.165e-14; nextEMgain4.494e-6. Native singletonbrightness comparator restored under additive pre-reference amendment without model/refit changes. Independent raw-reference/geometry review checks132audits,32/84contexts,26/77targets,allbudgets/20controls within7.816e-11arcsec. Allworkers terminal,94scientificbindings accepted.
Fourunlabelled families converge after92updates on50D1/D2groups;88of340groups have spectral embeddings. FrozenOrion10%ranking proposes14candidates,8with overlappingHOPS/VISION youthcontext and6without. RequestedHOPSrecovery8/26=30.77%,VISION8/77=10.39%; brightness retrieves2HOPS/2VISION at14reviews. Class0IDs81and91are retrieved, gaining7HOPSidentities overoldopticalHOPS71; requestedtargetunion8, not physicalstars. Native measurements/model/inputhashes, all132coregroups and32/84contexts retain singletons/ambiguities/misses.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| frozen infrared candidate groups | 14 | — | — |
| requested HOPS candidate recovery | 8 | — | 26 |
| Class0candidate recovery | 2 | — | 8 |
| Orion spectral-supported groups | 18 | — | 132 |
| confirmed physical births | 0 | — | — |
Next: Improve explicit uncertain/multi-band association and a native singleton ranking tier, compare actual observed-band colours and brightness, and develop broader unlabelled IR/dust transfer with representative starless/contaminant truth. Preserve optical/Gaia ordinary stars and compatible dust/youth/physicalownership/membership/temporalbirth-region scope. Publish E136-E153/localreviews afterMacunlock/readback; fullgoalactive,E121zero,12MiBstage/5GiBreserve.
Only18of132Oriongroups have coherent two-band scores; nineareambiguous. SixofeightClass0references are structurallyunscored:4ambiguous,2single-band despiteprevious100micronavailability. Commonfour-methodcohort hasonly2groups andno superiority; actual70/100or100/160colour baselines remain. Hidden45-cell predictionRMSE1.049 loses to brightness-offset0.603, especiallyOrion1.096vs0.564; tinyD3gain7cells doesnotvalidatebroadtransfer. At50%brightnessretrieves13HOPSvsfamilies9. Sparseconsumedrefs, upstreamsupervisedquality, smalltraining/localoptima, simulationestimatederrors andstaticbeamownership prevent populationpurity, physicalclasses/membership/age/birthclaims.