Matched source-screen baselines fail; per-band acquisition unavailable
6 Oct 2026 · embeddings
Does joint learning improve on simple missing-view fallback, and can consistent native per-band support address rejected optical measurements?
Retrieve native flags/calibration/composite fluxes and radii from36 fixed public sky stripes, with exact shared-measurement checks before applying the new consistent PSF guard. After a first503 service failure, preserve the prefix/35unrequested states and benchmark unchanged raw/learned optical and raw Gaia OR/AND screens on the same cached optical population without adopting a physical union.
Two per-band analytic/synthetic controls pass, including large integer masks, type/reference poisoning, signed colours, masked bands and independent finite-difference covariance. Native acquisition receives oneHTTP503/107-byte service response and zero new measurements; no retry. Independent cached replay checks4,013 optical/807Gaia native rows,35,487 Boolean decisions, all identities and nine method metrics. Accepted comparison workers terminal0; failed acquisition and pre-registration missing-import logs remain.39science bindings preserved.
Raw optical OR Gaia retrieves791 point proxies versus raw758, with unchanged228 galaxy proxies and83unknowns. Raw Gaia alone gives388points/two galaxies/38unknowns at99.49% conditional precision on the same optical universe. The new per-band guard/feature implementation is ready but remains unapplied until native inputs are available. All original scored candidates,63 images,198/729 reviews and dust/molecular evidence remain unchanged.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| source population entries | 3943 | — | — |
| raw fallback point proxies | 791 | — | 1030 |
| learned fallback point proxies | 843 | — | 1030 |
| accepted new native measurement rows | 0 | — | — |
| unrequested native queries | 35 | — | — |
Next: Use cached magnitude/availability-conditioned family and readout ablations against matched raw missing-view baselines. When meaningful service evidence permits a new bounded registration, obtain native per-band support and composite measurements before fitting changed features. Freeze against representative development calibration and validate a distinct source-selected field. Retain compatible dust/embedded and physical youth/membership/temporal-birth scope. Publish E136-E141 after manual Macunlock/readback; full goal active, E121 requests zero.
Every one of nine90%P/80%R gates fails. Raw-OR precision/recall is77.63%/76.80%; learned-OR64.16%/81.84% with843point proxies. E140 joint loses18point proxies and removes33galaxy proxies net versus learned-OR; against raw-OR it adds34points but210galaxies. Native catalogue type6 gives75.06%/87.67%; adding clean gives84.45%/74.85%. Legacy clean selection inherits pipeline morphology dependence. No new native measurement support, fit, calibrated union, first-use validation, physical truth or birth is demonstrated; unrequested regions remain unknown.