Fresh regional transfer and a 729-source optical review list
5 Oct 2026 · embeddings
Do the frozen unlabelled stellar scores retain precision, recovery and useful review ranking on the distinct E131 cohort?
Use unchanged E127/E130 models and E113 operating thresholds; all 14570 optical sources, 2776 Gaia sources and 232 spectra. Evaluate both-direction associations, unknowns, original and padded requested targets, raw baselines and fixed 1/5/10/20/50% review budgets without fitting or threshold selection.
Nine analytic controls pass. Independent 1878980 optical/spectral, 299016 Gaia/spectral and 21481112 optical/Gaia decisions; 378 reference alternatives, 2648 optical/Gaia alternatives, 87 bounded moving-pair minima. All 911 curve rows, 250 budgets, 1000 requested budgets, 75 matched comparisons and candidate exports replay; science workers terminal exit0. Independently checked exact 729-source shortlist.
Fresh regional conditional optical gate passes: 41 STAR/1 QSO selected, 97.62% precision and 100% recall; raw shape selects the same qualified identities. At 5% (729 reviews), latent shape retrieves 35/41 STAR and no qualified nonstars, versus raw 14 STAR/1 nonstar and brightness 35/11. Learned gains22/loses1 STAR relative to raw. Astrometric refinement and standalone Gaia recover37/41 with0 nonstars, identical labelled retrieval to raw. Original-field quality recovery is optical40/41 and Gaia36/41; 5% optical latent34/41 versus raw13/41. Linked 42 existing spectral STAR proxies and all prior stellar/dust/103 molecular tiers retained.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| fresh conditional optical stellar recovery | 41 | — | 41 |
| qualified optical nonstellar selections | 1 | — | — |
| optical operating source entries | 5422 | — | 14570 |
| optical review entries at fixed five percent budget | 729 | — | 14570 |
| qualified STAR references in five percent review | 35 | — | 41 |
| review entries without qualifying spectral truth | 694 | — | 729 |
Next: Measure candidate population purity with a source-selected review sample rather than target-selected spectra; validate transfer at faint, crowded and embedded populations. Retain permissive optical and refined astrometric tiers, improve learning only against matched raw baselines, and continue compatible dust/infrared/molecular and birth-region refinement. E131 is now consumed development for future model changes; new choices need another fresh cohort. Full goal active, E121 upload still pending with zero requests.
Conditional reference coverage is only59/14570 optical and61/2776 Gaia. 694 of729 review entries lack qualifying spectral truth; zero observed nonstars is not shortlist purity. B supplies29/41 optical positives, D none. Nominal 5% recall Wilson interval71.56-93.12%; regional dependence and target selection remain. Learned full optical operating burden5422 exceeds raw3978 with the same reference retrieval; Gaia learning has no demonstrated labelled advantage. E128 precision failure remains valid. Missing/ambiguous and all raw requested targets remain; physical catalogue overlap prohibits summing counts. No all-sky purity/completeness, calibrated ownership/union, youth, membership, dust class or birth established.