Amplitude-independent spectral shape yields a limited diagnostic gain but fails primary retrieval
6 Oct 2026 · embeddings
Can unlabelled native infrared shape families remove the brightness shortcut and improve candidate retrieval while preserving singleton follow-up?
Before fitting register log-flux contrasts, shared100micron covariance, explicit oblique missing-band operators, four linear-Gaussian families on the same50unlabelled D1/D2groups, seed1550 and covariance floor0.03. ExcludeD3/Orion/classes from fitting; freeze model,717anchors/1336views,18review primary and original173/43/483cohorts before consumed-reference readout. Register a separate alternating shape/singleton-brightness diagnostic, exact common-input comparators and full cost curves.
Seven controls pass. Independent native replay checks340groups/642cells exactly; scalar initialization/finalEM replay checks968embedding cells within2.443e-15 and9hidden-contrast cells within2.221e-16. Independent application replay checks15123quantities within7.106e-15. Raw-reference replay checks305466VISION primitives/95402nulls,36956HOPS primitives,5051982geometry admissions within8.995e-11arcsec,483audits,32/84contexts,all class/common-cohort/budget unions and3503curve points. All workers observed terminal;71scientificbindings accepted.
Four shape families converge in64updates.283of717anchors have actual multiband shape; reliable core68of173. Primary18retrieves5/26HOPS and5/77VISION, tied with brightness, but0/8Class0 versus brightness3/8. Ten overlapping youth-context anchors represent onlyfiveHOPSIDs, with8unknownanchors. Separate100micron18review diagnostic retrieves14HOPS/14VISION/5Class0 versus brightness13/13/6; on identical26shape-supported inputs brightness13HOPS/5Class0. Preserve all18primary,18diagnostic,18hybrid and earlier candidates.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| shape-supported native anchors | 283 | — | 717 |
| reliable core shape support | 68 | — | 173 |
| primary HOPS target recovery | 5 | — | 26 |
| primary Class0 target recovery | 0 | — | 8 |
| diagnostic100 HOPS target recovery | 14 | — | 26 |
| confirmed physical births | 0 | — | — |
Next: Retain multiple spectral phenotypes and cheap measured-colour/100micron/singleton follow-up, improve short-infrared and shape evidence, and broaden unlabelled development plus representative contaminant/starless truth before fresh regional transfer. Do not retune the failed primary from these labels. Preserve ordinary optical/Gaia stars, dust/embedded candidates and physical membership/temporal birth-region scope. Fullgoalactive,12MiBstage/5GiBreserve; zero astronomydownloads/images/E121requests/Gitrewrite.
Primary gain fails and no consistent advantage across budgets. At35reviews apparent10versus brightness9HOPS becomes10versus10on identical shape inputs, with fewerClass0. Shape support caps at16/26HOPS/6of8Class0;105reliable-core singletons remain unscored for shape. The predeclared hybrid retrieves only3HOPS/0Class0at18reviews. Diagnostic measured70/100colour retrieves14HOPS/6Class0atlower17cost. PredictiveRMSE0.411versus0.710uses only3sources/9correlatedcells; two small families are floor-limited. Sparse consumed references, targeted exposures, upstream supervised quality and missing negative truth leave populationprecision null; no physical classes, ownership, membership, ages or births adopted.