Complementary retrieval restores 33 censored molecular follow-ups but brightness wins
5 Oct 2026 · embeddings
Can class-blind bright-cloud and learned warm-source branches recover missed candidates at the same review budget?
Reuse frozen E134 embeddings. Primary policy alternates native brightness and warmth, skips duplicates and retains unsupported embeddings through the raw branch. Compare native/latent brightness, unchanged warmth, raw dual and temperature, with fixed full-region budgets and exact common support. No labelled quotas or refitting; freeze orders before reference evaluation.
Three controls pass. Independent Decimal/deque reconstruction checks all 189,208 test/calibration records, 1,324,456 priority ranks, 795,144 observable/latent scores, 350 budget comparisons and 1,138,970 reviewed memberships. Maximum score difference 1.78e-15. All calibration fractions, gained/lost identities and 10,140 molecular-view cells replay. Accepted workers terminal exit0; science seal has 35 bindings.
At 4,730 reviews the primary selects 338 reliable molecular links, including 33 native 70-micron-upper-limit clouds; unchanged warmth retrieves 285 links and six limit cases. Relative to warmth, primary gains 60 and loses seven links, with 278 in common; all 189 temperature-selected links are retained. The 338-cloud native measurement view retains actual positions, velocities, line flags/intensities/errors and model/input hashes; 31 overlap the prior 103-source refinement queue. All 94,604 requested clouds and prior stellar/dust/line tiers remain.
| Metric | Value | 95% CI | n |
|---|---|---|---|
| full requested cloud entries | 94604 | — | — |
| fixed five percent review entries | 4730 | — | — |
| reliable molecular review entries | 338 | — | 433 |
| review entries with native seventy micron limits | 33 | — | — |
Next: Audit or acquire independent stellar/YSO classification and spatial/velocity/foreground ownership for actual candidate sources, with source-selected purity and faint/crowded transfer. Retain brightness as the strongest cheap benchmark and the useful unsupervised representation; stop treating another quota on the same references as fresh validation. Preserve ordinary-star, compatible dust/embedded and birth-region scope. Full goal active; E121 pending, zero requests.
Primary loses to native brightness at every budget: at 5%, 338 versus 340, gaining two and losing four reliable links; on common support 316 versus brightness 326. Latent brightness adds only two links at 20%, raw dual one at 50%, with no uniform advantage. Every joint 90% precision/80% stage-recall gate fails; pooled dependent-stage precision 88.05%, recall 40.80%. Stage reuses 70-micron input; molecular activity is targeted consumed context, not stellar purity or recall. Upper limits do not prove a cold or starless physical class. Missing errors, approximate likelihood, blending and ownership remain. No new star, birth, physical union or robust all-sky detector is established.