Signed-flux inputs restore numerical support but expose severe region-dependent contamination
6 Oct 2026 · embeddings
Can cached native signed fluxes preserve weak-band evidence without inventing censored limits or silently accepting contaminated measurements?
Freeze all 1,967 sources and numeric signed flux/error availability separately from magnitude support. Retain zero/negative/flagged/extended/saturated measurements. Register a conservative diagnostic requiring no contamination flag, zero saturation, 0<profile-fit chi-square<=3 and extended flag zero. Do not adopt a limit without native ph_qual, change candidate ranks, inspect references or fit a model.
Fourteen controls pass. Independent six-response XML/Decimal replay checks 88,515 native primitives including16,586 nulls,125,888 per-band diagnostics, every exact native ID/row/field/role/hash, support and quality state, source count and field aggregate. Maximum numeric difference4.021e-14. All workers observed terminal0;28 scientific bindings accepted.
Numeric signed flux/error pairs give at least two WISE bands to1,966 of1,967 sources versus1,685 magnitude-pair sources. Preserve3,201 additional numeric band pairs and1,539 nonpositive fluxes in native count units. All source measurements and explicit unknown censoring states remain; these are input support, not recovered stars, recall or valid Gaussian errors. Conservative two-band quality diagnostic retains1,244 sources overall.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| sources with two numeric signed flux pairs | 1966 | — | 1967 |
| sources with two conservative-quality bands, diagnostic only | 1244 | — | 1967 |
| Orion sources with two conservative-quality bands, diagnostic only | 8 | — | 302 |
| additional numeric band pairs | 3201 | — | — |
| nonpositive numeric fluxes retained | 1539 | — | — |
| confirmed physical births | 0 | — | — |
Next: Specify signed-flux likelihoods with explicit contaminated/unknown measurement states, or acquire bounded native quality/deblending metadata before censoring adoption. Compare simple noise-aware baselines and current magnitude ranks at equal review cost/common support without erasing flagged-source and cold far-infrared follow-up. Freeze model and allocation before reference inspection; obtain representative independent calibration and fresh transfer. Full stellar/dust/physical-temporal birth-region goal active. Zero astronomy requests/E121/fits,12MiB cap/5GiB reserve; external update queued behind pending separate readback approval.
Quality varies sharply with geography: conservative two-band support is only8/238 D1 and8/302 Orion, versus352/353 D3,295/296 D4 and408/410 D5. Neither D1 nor Orion has a qualifying W3 orW4 band. Orion carries251 possible-spurious W3 flags and266 W4 flags; flags do not prove every source false. This diagnostic is not adopted as a selection. Missing ph_qual/deblending/epochs and unknown sampling/error correlations prevent valid censoring or naive clean Gaussian flux adoption. No model, readout, population precision/recall, physical classes, ownership, membership or birth established.