Dust-region screening shortlist exposes failed spatial precision–recall transfer

Can frozen unsupervised dust representations support useful calibrated region retrieval?

Class-free buffered splits within five held-out regions; calibration-only90% precision/80% nominal Wilson lower and at least20references; freeze before evaluation, compare raw scores and requested80%recall.

Independent native availability, scalar grouped-threshold and table review checks94604 rows,24687 candidates,473020 feature states and79568 score groups; all split/calibration/Wilson/budget/denominator decisions. Original strict-numeric reader refusal retained with separately bound numerical-contract amendment.

24687 research-region candidates including1179 unknown-stage rows. Evaluation precision77.01–94.80%, recall44.31–87.34%; every regional joint gate fails. Pooled precision84.81%, recall58.76% versus raw84.83%/59.24%; no embedding gain. No downloads/refit; compressed stage under9MiB.

Next: Improve label-free nonlinear/missing-view ranking with raw baselines and independent references; freeze before fresh regions. Use native shortlist for bounded spectral/image/temporal candidate refinement; complete robust star/star-birth-region goal stays active.

Unsupervised PCA, supervised stage probe; catalogue references and OOD heuristic, not physical truth. All labels previously consumed for development, not fresh validation. Unbound is unknown, shortlist includes calibration/buffer and supplies no deployment purity. No class/membership/birth adopted.