Stress-Testing DANTE under Detector Domain Shift: a Representation-Coherent Reanalysis of LIGO O4a

Abstract

This sixth version of the Domain-Adaptive Network for Transient Evaluation (DANTE) preprint stress-tests an unsupervised transient-noise pipeline under representation mismatch and observing-run adaptation. We reanalyse 10,429 detector-time strain candidates from 42 LIGO O4a sessions using frozen DINOv2 patch embeddings and a Top-k multiple-instance score. Candidate and native-background Q-transforms share Q in [4,64], and detector-specific thresholds are calibrated from 5,000 run-native windows by temporal-block bootstrap. The coherent analysis yields 6,365 ROBUST, 1,275 AMBIGUOUS, and 2,789 BACKGROUND statistical dispositions; 4,676 of 10,372 paired historical dispositions differ from the cross-representation v5 analysis. Direct controls resolve an O3b-O4a score shift and reduction after native adaptation for H1, but not L1, while known-glitch separation is detector- and morphology-dependent. Replicated studies quantify population-dependent sensitivity to background draw, clustering seed, dictionary size, and whitening, demonstrate native-index absorption, and identify conditional low-Q blindness. A conservative H1-L1 max-shift screen yields 13/8,806 values above threshold, but its on-source values and pooled per-event null maxima are not exchangeable. The primary two-null PEM endpoint shows no resolved ROBUST-BACKGROUND enrichment (p=1.000), and two catalogue overlaps are consistent with a circular-shift coverage proxy (p=0.651). Simulation-only compact-binary controls show detector- and distance-dependent disagreement between novelty, native disposition, and physical coincidence. We withdraw the v5 discovery, rate-limit, catalogue-recall, and survey-wide stability interpretations. The supported result is a measured set of failure modes and validity conditions for unsupervised detector characterization, not a new glitch class or an astrophysical search.