Supports· moderateTRN-E015
ML artifact-scoring: ensemble effects persist among likelier-real candidates
- what the source states
- A model trained on 250 transient image pairs classified as real versus plate defect by expert visual review (out-of-fold AUC = 0.81; sensitivity and specificity 0.71) assigned each of the 107,875 catalog transients a probability of being real. After controlling for ML-identified artifacts, nuclear-window counts remained significantly elevated (p = .024), the highest-probability transients were more likely to fall in nuclear windows (p < .0001), and the shadow deficit remained significant (p < .001) and was largest among highest-probability transients (p = .003).
- what we infer
- A real internal-robustness result: the obvious visually recognizable defect classes do not trivially explain the ensemble signals. It is not external validation — the training labels are expert judgments about the same kind of imagery under dispute, not physical ground truth, and the whole analysis stays inside the contested catalog lineage.
- limitations
- Training labels derive from expert visual classification, not from independently known photons-versus-defects ground truth.
- Same 107,875-object lineage as the original claims; author overlap with both the original papers and the Doherty replication.