Bistro Research
Partial

Broader stellar families predict better but lose candidate recall

5 Oct 2026 · embeddings

Imported record: produced before this platform existed
Question

Does broader native label-free density learning improve stellar retrieval beyond the original selected training population?

Hypothesis

Two registered eight-family models: unfiltered-patch training only and augmented rare-colour training. Preserve all 9230 native identities, native measurement gates and original role hashes; exclude calibration/test observations from fitting. Calibration-only interpretation and fixed raw baselines.

Test

Independent literal features and inverse/determinant Gaussian replay checks 73840 feature states, 166140 scalars, 12948 withheld values, 3570 reference decisions, 60 budgets and 4569 experimental candidate records. Maximum posterior difference 3.031e-14. Existing streaming CLI checks both models on 61 sources with two chunk sizes.

Finding

Fits converge in 359/153 updates on 2428/3334 sources. Original qualified test precision/recall: broad-only 97.12%/68.71%, augmented 97.50%/79.59%, E112 94.29%/89.80%. Nontraining E113 stars retrieved 11/25,13/25,19/25; no other selected there. Augmented withheld RMSE 0.449 versus E112 0.504, an 11.03% improvement; predictive gain does not imply retrieval gain. Experimental lists 2242/2327; original 3225 retained. No downloads.

Next: Preserve E112 recall anchor and both experimental models; investigate rare-family/missing-view discrimination with independent context and contaminant coverage, without test-based winner selection. Freeze before fresh surveys; retain compatible dust/103 molecular candidates and temporal refinement. E121 upload permission remains pending. Full robust goal stays active.

Caveat · open

Density fit is label-free but classifier interpretation is reference calibrated. Original labels and E113 are consumed development evidence, with source conditioning, tiny reference coverage and missing views; no fresh global precision. Native uncertainties gate values but covariance/censoring are not modeled. Original log seal captured pre-final-print hash; exact additive correction and first newline-reader refusal retained. No physical class, membership or birth adopted.