Bistro Research
Partial

Fresh optical transfer improves small-budget recovery, while high-recall precision fails

5 Oct 2026 · embeddings

Imported record: produced before this platform existed
Question

Does the unchanged unlabelled optical embedding transfer to new public regions with useful precision, recall and review burden?

Hypothesis

Freeze E127 model/scores/thresholds before eight nonoverlapping RA-shifted regions,64public optical and8spectral GETs. Keep native unknowns/alternatives, raw baselines and requested denominators. Require>=20positive proxies and90%precision/80%recall without retuning.

Test

Independent754160 native primitives,224628features/1347768covariances,37438poison controls and zero prior identity overlaps. Inverse inference224628scores/299504posteriors, max1.261e-12/1.024e-13. Haversine6524666decisions/789alternatives/675audits,max3.808e-11arcsec;775curves,30budgets,120requested-budget rows,14654candidate records. Separate network-context correction preserves initial DNS failure beforeHTTP and all72exactquery bytes.

Finding

37438new optical identities,675spectra,155qualified(65STAR/90other);23362supported/14076unknown,3.91MiBresponses. Primary63STAR+13others(11QSO/2GALAXY),82.89%precision/96.92%recall; same76reference identities as raw. Gate fails. Primary14565/dense14224/raw11438 candidates. Fixed5%budget1872: primary49STAR+2QSO(96.08%/75.38%) versus raw30+5(85.71%/46.15%); dense18+7 fails. Requested quality field operating60/65unique,61/65any;5%48/65versus29/65raw. Export63retrieved stellar proxies with native provenance.

Next: Retain the verified1872-source primary review shortlist and all permissive stellar/dust/103molecular tiers. Refine compatible native optical/IR/astrometric evidence for QSO and failed stellar targets, learn more star-specific structure without class loss, freeze before another fresh cohort. Dust/embedded and molecular refinement remain required; E121upload pending permission, full goal active.

Caveat · addressed_unverified

Spectral proxy coverage0.414%, target-selected quality labels, nearby fixed regions and approximate noise limit generalization. Conditional65positive denominator excludes one ambiguity/three unsupported field targets and includes padding. High-recall precision still fails; no calibrated union, new-star discovery, physical class or birth. E128 becomes consumed after assessment; no test-selected ranking or threshold swap.