Censored embeddings improve spectral prediction and retain 285 molecular follow-ups
5 Oct 2026 · embeddings
Do native measurement errors and 70-micron limits improve label-free far-infrared representation and candidate retrieval?
Five spatial four-component Gaussian families use native asinh fluxes, propagated errors and a single censored coordinate. Training uses 12,000 class-blind sampled sources per fold, excluding test/calibration regions and one-degree buffers. No stage or molecular evidence enters fitting or prototype warmth ranking. Freeze before evaluation and compare fixed full-region budgets with raw colour, temperature, brightness and E120.
All five models converge in 141/150/200/75/123 updates. Independent checks cover 94,604 native sources, 189,208 test/calibration records, 1,830,004 posterior/rank/latent scalars and all 300 budget comparisons. Maximum difference 1.47e-13. The preregistered hidden-band check compares 301,599 measured cells and 603,198 independent mean/density scalars. Five analytic tests pass; accepted terminal science seal has 64 bindings.
Model support is 83,182 clouds; all 11,422 unsupported sources remain in requested denominators. Hidden 160/250/350/500 micron RMSE is 0.527/0.283/0.224/0.230, versus source-offset 0.730/0.350/0.379/0.467. At 5% (4,730 reviews), the new families retrieve 285 reliable molecular links versus old E120 171, temperature 189 and brightness 340; exact common-input counts are 269 versus brightness 288. A traceable 285-cloud molecular review view contains actual coordinates, velocities and line measurements: 141 overlap the old 5% list, 144 are newly reviewed relative to it, and 15 overlap the prior 103-source queue. All prior stellar/dust tiers remain linked.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| full requested cloud entries | 94604 | — | — |
| fixed five percent cloud review entries | 4730 | — | — |
| reliable molecular links at five percent review | 285 | — | 433 |
Next: Test complementary cold/bright and warm retrieval against the strongest brightness and temperature baselines, retaining independent molecular/stellar refinement and all requested sources. Measure source-selected stellar purity and faint/crowded/embedded transfer; resolve ownership, youth and cloud membership before birth claims. Do not tune physical classification on dependent 70-derived stages. Full goal active; E121 remains pending with zero requests.
Every joint 90% precision/80% stage-recall spatial gate fails: pooled dependent-stage precision 83.44%, recall 62.42%. The 5% stage result (4,718 positives/12 negatives) is dependent because 70-micron detections define the catalogue stage; raw70 colour selects 4,730 positive stages. Brightness retrieves more molecular links at every budget, and learned ranking loses to old families/temperature at 50%. Molecular activity is previously consumed targeted context, not stellar truth, membership or birth. Asinh errors, limit/5 sensitivity and independent-band noise are approximations; blending and missingness remain. The original heavily censored synthetic recovery did not converge and is preserved. No new star, calibrated physical class, all-sky purity or robust cold-source recall is established.