Joint unsupervised families restore recall but optical-only precision fails
6 Oct 2026 · embeddings
Can star-specific optical/astrometric families retain low-motion and missing-view sources without the hard motion conjunction?
Fit eight constrained nine-coordinate error-deconvolved families on class-blind cached development inputs: point-shape blocks for four families and stationary astrometry for four alternatives. Marginalize missing views, withhold reference-near identities and freeze before E113 operating calibration. Keep full COSMOS denominators, raw/brightness/hard-conjunction comparisons and no first-use validation claim.
5,000 unlabelled rows fit in 172 monotone updates; 974 have both views and 99 astrometry alone. Of 36,611 E126 rows, 563 actual identities are withheld; 897 is the broader exclusion-ID pool, including 334 outside this universe. Three analytic/synthetic controls pass. Independent checks cover every role, training feature, a constrained EM update, 286,720 score/embedding entries, 2,295 PR rows, five budgets, 3,943 source reviews, 1,380 candidates and 2,170 native requested HST rows. Maximum score difference4.66e-10; final mean likelihood matches exactly. All accepted workers terminal0. Historical failed native text comparison is preserved, numerical display replay passes. Original243 bindings plus five additive role-scope bindings retained.
The new model retrieves 825 matched point proxies at 80.10% recall, compared with parent811/78.74% and strict refinement354/34.37%. It gains33 point IDs and loses19 versus parent, net+14. Both-view support gives425 point/two galaxy/32unknown selections and99.53%/97.93% conditional P/R on a smaller434-point denominator. Astrometry-only support recovers26 point proxies/eight unknowns. All1,380 provisional candidates retain native source evidence and joint embeddings; old198/63-image/729-source and dust collections remain.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| source population entries | 3943 | — | — |
| joint candidate entries | 1380 | — | — |
| joint point proxies | 825 | — | 1030 |
| joint galaxy proxies | 438 | — | — |
| deeper requested recovery | 825 | — | 2170 |
Next: Improve native flag/repeat/pixel measurement support and optical-only extended-source discrimination; use a matched raw joint-view comparator and representative development calibration, then freeze before a distinct source-selected validation field. Preserve compatible dust/embedded and physical youth/membership/temporal-birth refinement. Publish corrected E136-E140 after manual Macunlock and exact readback. Full goal active; E121 requests zero.
Full precision is65.32% with438 galaxy proxies/117unknowns; whole-selection morphological purity bounds59.78–68.26%. Every full gate fails. Optical-only precision is46.17%. Of168 point proxies without inference,161 fail the optical clean flag and seven the anchor SNR, all lacking usable Gaia. At198 reviews joint retrieves172 points versus old188/raw185; at789 it gives680/37galaxies/72unknowns versus old690/50/49. Best descriptive recall at>=90%P is71.46%, so no current score threshold satisfies both gates; no posthoc threshold adopted. Bright performance does not extend to faint/missing/crowded inputs. Requested recovery825/2,170 differs from matched recall; sparse E113/E125 spectral calibration looks perfect but fails on broader morphology. No physical probability, dust transfer, new star, member or birth confirmed.