Immediate goal: diagnose information and generalization before expanding the search

Human-authorized reformulation, 14 October 2026. The physical target and long-term requirements in RESEARCH_GOAL.md remain unchanged: embedded protostars in Hi-GAL catalogue-prestellar clumps, ultimately independently validated at survey scale.

The next bounded deliverable is a retrospective diagnostic ablation, not another candidate discovery claim. Determine whether ranking consumed chemistry proxies comes from physical covariates, survey availability, actual molecular information, or the learned representation. Compare training and spatially held-out performance, fixed representation capacities and regularization, and nested label-count curves. Report every registered contrast, unknown and failed gate. Do not select a model from this study for deployment or call these reused outcomes fresh validation.

The representation is unsupervised; the downstream proxy ranker is supervised. Dust emission and an embedded star can coexist. Survey association, a chemistry flag, an embedding score and a stellar birth are different claims.

After the diagnostic, specify the smallest independent evidence requirement that could change the conclusion. Prioritize association quality, complementary view coverage, physical evidence and an untouched evaluation cohort over a larger architecture. The current 27 consumed positive proxies cannot establish population precision/recall; zero has the registered temporal view. No additional acquisition, threshold change, archive restart or model promotion is authorized by this document. The human requested this study after seeing the original outcomes; freezing its contrasts before execution does not make it an outcome-blind preregistration.

Implementation and limits: fixed diagnostic protocol.