Why a cutting-edge AI pipeline missed every gravitational-wave glitch it was hunting

A blind injection test of eight synthetic glitch types into real LIGO O4a data exposed a fundamental flaw: the DINOv2 vision model compresses each spectrogram into a single average value, so any glitch occupying less than 5% of the image gets washed out entirely. Tighten the detection threshold to control false alarms and recall drops to zero — even for signals 430 times above the noise floor. The culprit is global average pooling, and the fix points toward patch-level scoring in next-generation pipelines.