Native PS1 compact-source representation learns from separate unlabelled regions
6 Oct 2026 · embeddings
Can a separately trained, noise-aware PS1 representation preserve compact-source evidence without stellar labels?
Learn four compact and four flexible families from native colours, shapes and brightness, allowing a shared empirical Kron/PSF offset. Keep a geographically separate unlabelled diagnostic region and every missing input.
Six counted unfiltered regional GETs yield4454stack rows/3356objects. Independent replay checks472124numeric primitives and4454positions, with no COSMOS ID overlap. Three analytic controls pass; scalar replay checks31178coordinates,124712covariance entries,22270quality decisions and3356representatives. Independent inverse-Gaussian/final-EM replay matches154376embedding values within2.14e-13;52closed scientific bindings accepted.
The model converges after225updates on951unlabelled D1/D2objects, learning native compact shape centre(-0.07689534,-0.06429605). All3356objects retain native source/stack/epoch/hash/role evidence.1577have supported scores, including643of1389D3objects;1779remain unscored. No catalogue/reference labels, COSMOS fitting/calibration, threshold changes or images enter the work.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| native development objects | 3356 | — | — |
| unlabelled fit objects | 951 | — | — |
| supported shape embeddings | 1577 | — | 3356 |
| EM updates before convergence | 225 | — | — |
| confirmed physical births | 0 | — | — |
Next: Obtain representative development reference coverage outside consumed COSMOS, compare learned/raw morphology/brightness on identical sources, freeze calibration-only readouts before a distinct source-selected validation field, and retain all-requested unknown/ambiguous denominators. Preserve dust/embedded and physical birth-region scope; publish pending findings after unlock and exact readback.
These are compact-source embeddings, not calibrated stellar probabilities or detections. Shape availability is46.99%ofthis selected catalogue, not recall. D3has no representative labels and supplies no accuracy claim. Three small northern regions, missing views, PSF/Kron errors, the fixed shape width and local-optimum identifiability limit transfer. Unresolved galaxies/quasars, blends, dust/embedded transfer, physical youth/membership and births remain unresolved.