A night watchman doesn't know which song a burglar will sing, but he hears any sound and calls for backup. That's roughly how the new gravitational-wave search works: it doesn't keep ready-made "notes" for every source in mind, but catches any burst in the fabric of space.

The LIGO detectors and their neighbors usually look for signals with matched filtering: data are compared against millions of pre-computed templates, as if searching for a familiar melody in the noise. But a supernova flash or the shudder of a neutron star has no exact score, so they need a different kind of hearing. A wavelet filter splits the data stream into short waves of different durations and looks for the ones that stand out from the noise. It's like how an ear picks out individual drumbeats in a general rumble.

When suspicious fragments pile up in one detector, they are combined into an event. Then events from different detectors form a graph: the nodes are the finds, and the edges are pairs that could have come from a single source. Such pairs used to be assessed with a simple formula, but now a neural network looks at the graph. It is trained on examples and decides for itself which coincidences are real and which are just random noise.

A test on real data showed that the network finds almost all black hole and neutron star mergers, and for supernova flashes, more than half. Once it spotted a signal in the stream that turned out to be event GW250114: the loudest candidate, even though no ready-made template had been slipped in. So the watchman, not knowing the song, still raised the alarm.

Previously it was unclear how such a search could work in real time and not only find but also reconstruct the signal. Now it's clear: wavelets catch the burst, and the neural network checks whether a single source is playing in two detectors. This pipeline can already run at existing observatories and will help hear what no one expected.