Native astrometry supports candidates but strict refinement loses recall
6 Oct 2026 · embeddings
Can unchanged unsupervised astrometric families refine the whole COSMOS population and recover optical screening misses without hiding missing evidence?
Acquire all Gaia sources in four fixed public stripes, score the unchanged E130 model/thresholds, and associate to native optical epochs and the documented 2003–2005 HST interval. Preserve every parent source, fixed budget, unknown and raw comparator; do not tune on consumed COSMOS labels or adopt a new union.
Four native GETs return 807 Gaia rows; 13,719 native cells and every score/posterior/latent value replay. Independent checks cover 6,476,982 optical/Gaia and 188,466,780 HST/Gaia decisions, 846/5,419 complete alternatives, 4,190 bounded minima, 8,621 metric rows, all 4,013 optical associations/658 Gaia audits and all 198 preserved review entries. Maximum score difference 9.32e-10 and separation difference 6.22e-11 arcsec. Twelve original controls and one corrected calendar control pass. Additive full-calendar replay uses [2003,2006], retaining the historical [2003,2005] run and failed control margin. All primary results remain unchanged; one new three-arcsecond alternative is not adopted. Original95 and corrected127 bindings retained; accepted workers terminal exit0.
The original 198-source review gains exact optional Gaia evidence: all have unique alternatives, 194 have supported selected refinement (184 point proxies, two galaxy proxies and eight prior unknowns). Four prior point proxies retain poor/absent fits. Learned optical refinement selects 384 entries with 354 point proxies, two galaxy proxies and 28 unknowns: 99.44% conditional precision. Gaia-only selects 430 entries with 389 point proxies/two galaxies/39 unknowns. Its point-proxy reference IDs recover 32 optical parent misses, including 25 unsupported optical inputs; native rows and separate universes remain traceable.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| optical source population entries | 3943 | — | — |
| Gaia source population entries | 658 | — | — |
| supported original review entries | 194 | — | 198 |
| refined optical point proxies | 354 | — | 1030 |
| Gaia only point proxies | 389 | — | 548 |
Next: Improve unsupervised joint optical/astrometric inference for faint, low-motion and missing views against raw baselines, calibrating only on development and freezing before a distinct source-selected field. Inspect morphology/astrometry conflicts with independent physical evidence. Preserve ordinary-star, embedded/dust and molecular/birth-region tiers; publish corrected pending phases and native reviews after Macunlock and exact live readback. Full goal active; E121 requests remain zero.
Every joint 90% precision/80% recall gate fails. Optical refinement loses 457 parent point proxies and reaches only 34.37% matched optical recall; Gaia-only reaches 70.99% on its different 548-point-proxy denominator, and 389/2,170=17.93% of deeper requested HST proxies. Learned optical refinement loses one more point proxy than raw; Gaia learning gains two and loses one, only net+1. Astrometric ordering at 198 reviews retrieves 176 point proxies versus parent 188. Both galaxy labels conflict with strong native motion and remain unchanged pending physical review. Missing/weak Gaia is not nonstellar; masks, epoch/rounding, point-source AGN and absent QUERY_STATUS limit scope. No physical union, new star, membership, dust transfer or birth is established.