Brightness-aware families improve ranking but lose faint-star recall
6 Oct 2026 · embeddings
Does native brightness context improve unsupervised stellar discrimination beyond missing-view fallback without sacrificing faint-source recall?
Append native log r-band PSF flux with its shared-anchor measurement covariance to the exact paired5,000 training inputs. Fit eight10D constrained families once; freeze full odds and joint-minus-context conditional evidence before E113-only operating calibration. Keep physical-support abstention and all reference denominators; no COSMOS fitting or threshold adoption from consumed curves.
The fit converges in142monotone updates. Three revised synthetic controls pass; the original nonconverging fixture is preserved. Independent native/Jacobian/inverse-Gaussian replay checks820,506cells and340,480embedding values (max difference6.99e-10), a constrained EM update and native fixed covariance widths. Independent calibration/evaluation checks162calibration rows,4,590curve rows,10budgets,3,943source reviews,67,031Gaia cells and2,170requested targets. All workers observed terminal. Corrected200-binding closed-log seal supersedes the historical premature finalizer-log digest; science unchanged.
At789reviews full odds retrieves703point proxies/24galaxies/62unknowns versus E140680/37/72 and optical690/50/49. Relative to E140 it gains47point identities and loses24, net+23, while removing13galaxies net. The fixed conditional threshold selects491sources with450point/two galaxy/39unknowns:99.56%qualified-reference precision. It retrieves62additional point IDs versus raw Gaia and loses none without adding galaxy-reference IDs. Both frozen candidate tables and494distinct source-evidence rows retain native coordinates/errors/epochs/model provenance; old63-image/198/729/1380and dust collections remain.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| source population entries | 3943 | — | — |
| conditional candidate entries | 491 | — | — |
| conditional point proxies | 450 | — | 1030 |
| conditional galaxy proxies | 2 | — | — |
| deeper requested recovery | 450 | — | 2170 |
Next: Improve native measurement support and use representative development calibration, then freeze before a distinct source-selected validation field. Stop threshold tuning on consumed COSMOS. Preserve permissive ordinary-star, embedded/dust and physical youth/membership/temporal birth-region tiers. Publish E136-E142 and exact native reviews after manual Macunlock/readback. Full goal active; zero new requests/images/installs or E121 retry.
Every full90%P/80%R gate fails: conditional matched recall43.69%, full40.39%. Best descriptive recall at>=90%P is73.11%full and71.36%conditional; no curve threshold passes both. Diagnostic thresholds are not adopted. Conditional recall is1.74%for optical-only,24.09%for HST19–22and0/116for22–25; all168unsupported point proxies remain unavailable. At198reviews both give172points versus optical188. Sparse bright spectral calibration misrepresents broad performance; inherited clean flags depend on pipeline morphology. COSMOS is consumed, model odds are not physical probabilities, and no all-sky purity, dust transfer, new star, youth, ownership or birth is demonstrated.