Astronomers measure distances in the universe using explosions of special stars — Type Ia supernovae. These flashes always have nearly the same brightness, like light bulbs of known wattage. Knowing how much light reached us, you can calculate how far away the explosion is. But the accuracy depends on how well we know this 'bulb wattage.'

Neural networks took two datasets: a map of galaxy distribution and a list of supernovae. They adjusted the brightness so that both datasets agreed. It turned out that if you force the network to respect a simple rule, that light isn't lost and distances can be measured in two ways, the results become noticeably cleaner. This rule is more reliable than any assumptions about dark energy.

But the data revealed a strange dip: at a certain distance (redshift 0.4–0.5), supernovae appear slightly dimmer than expected. This dip is visible in all four cosmological models and with different settings. It can't be explained by chance — the statistical significance is 2–3 sigma. It's unclear whether this is new physics or an unnoticed error. But the fact that the dip doesn't depend on the model makes it worth a closer look.

Perhaps future data from new telescopes will show whether this anomaly disappears or turns into a discovery. For now, the neural networks honestly say: there's something in the cosmos we don't yet understand.