Source-selected HST references expose broad optical contamination
6 Oct 2026 · embeddings
Does the unchanged unsupervised optical screen retain precision and recall against a broader source-selected morphology reference?
Freeze E127 scores and operating points on a fixed COSMOS region before any HST request. Evaluate all sources against unfiltered ACS detections with one-arcsecond both-direction unique matching, explicit faint/quality/ambiguous unknowns, matched raw and brightness baselines and fixed review budgets. No HST features enter model inference.
Four SDSS responses provide 4,013 fresh padded sources; sixteen native Gator responses provide 116,770 ACS detections. Independent review checks 88,286 optical and 817,390 HST cells, 24,078 scores/32,104 family posteriors, 31,386 spherical pairs, 11,373 metric rows and all 3,943 source audits/198 reviews. Maximum score difference 2.31e-14 and separation difference 1.22e-10 arcsec. Three optical controls pass. Three failed original transport stages preserved under registered amendments; accepted workers terminal exit0 and science seal has 121 bindings.
Qualified image references cover 3,254/3,943 sources: 1,030 point proxies and 2,224 galaxies. At 198 reviews, learned shape retrieves 188 point proxies/two galaxies/eight unknowns, versus raw 185/three/ten and brightness 175/four/nineteen. Learned shortlist morphological purity bounds are 94.95–98.99%. At 789 reviews, learned retrieves 690 point proxies versus raw 671, with fewer galaxies (50 versus 71). The full native 198-source shortlist and every source score/provenance/unknown state remain available.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| source population entries | 3943 | — | — |
| qualified matched point proxies | 1030 | — | — |
| qualified matched galaxy proxies | 2224 | — | — |
| primary operating point proxies | 811 | — | 1030 |
Next: Add independent source-specific astrometric/multiband refinement and validate a new faint/crowded population without choosing thresholds on these now-consumed references. Preserve permissive ordinary-star, embedded/dust and birth-region tiers. Publish corrected E136/E137/E138 findings and native reviews after Macunlock and live readback. Full goal remains active; E121 pending, zero requests.
Every original 90% precision/80% recall gate fails. Learned operating precision/recall is 63.26%/78.74% versus raw 76.88%/73.59%; learned gains 53 point proxies but selects 243 additional galaxy proxies. Even assigning every operating unknown as a star gives only 66.07% morphological purity. At 5%, learned gains 90 point proxies and loses 87 relative to raw; small net gains do not preserve identities. Only 811/2,170 deeper requested HST point proxies are recovered; 1,104 lack SDSS alternatives. F814W22–25 learned precision is 29.63%. HST morphology, masks, unresolved AGN, static epochs/rounding and absent QUERY_STATUS limit interpretation. No all-sky purity, physical stellar truth, dust transfer or birth is established.