Assembling the puzzle of the Universe, astronomers are used to every new detail fitting perfectly into the already completed picture. But sometimes there are pieces that seem cut from another jigsaw. Twenty thousand active galactic nuclei, and eleven percent of them the neural network confidently rejects as alien. What are they: rare beasts of the galactic zoo or instrument trickery?

For three and a half years, the VST telescope strung together images of a patch of sky in the constellation Sextans, collecting photometric light curves — graphs of flickering light from distant sources. Behind this twinkling lies physics: a black hole, its accretion disk, dust clouds. But how to distinguish a rare cosmic event from just a bad image? Enter the autoencoder — a neural network that learns to compress typical data into a compact code and then reconstruct it. An atypical object will be reproduced with noticeable errors. The magnitude of this error serves as a flag for anomaly.

The autoencoder worked not with raw observations, but with 37 features — measured characteristics of variability, color, amplitudes. That way astrophysicists embedded their knowledge directly into the neural network. The anomaly threshold was set at the 95th percentile of error: everything reconstructed worse than 95% of the sample fell into suspicious. Out of 20,647 sources, 272 such objects were found—11.18%. Among them, active nuclei absolutely dominated (242 against 16 galaxies and 14 stars), with unobscured nuclei flagged almost always (90.7%!), and obscured ones three times less often.

But how to tell which anomaly is due to real astrophysics and which is an artifact of observations? Here comes SHAP — a method that decomposes the contribution of each feature to the final score. Like a magnifying glass, it shows which corner of the 'puzzle piece' sticks out from the overall picture. For anomalous AGNs, the main culprits were variability parameters: excess variance, amplitude of random walk. For galaxies and stars, anomaly was dictated by completely different features. This allowed separating physically interesting cases (unusual structure of the dust torus, intermediate nucleus type) from failures of the measurement methodology.

Ahead lies the era of LSST, when a flood of photometric data will overwhelm astronomy. We cannot do without automatic detectors. This work showed that autoencoders with Explainable AI can become not just filters, but meaningful tools for finding the rarest stages of galactic evolution. The next step is testing on synthetic light curves and adding contextual information, such as redshift and galaxy morphology. Now astronomers have a tool that doesn't just discard irrelevant details, but illuminates the puzzle ready to form an unexpected picture.