Studying neutron stars — the ultra-dense remnants of exploded giants — astronomers seek clues in gravitational waves: ripples in spacetime born during collapse. Detectors like LIGO catch these tremors using ultra-precise laser measurement. Machine learning programs are supposed to infer the properties of the inner matter from the wave shape. But standard testing breeds a false sense of success — like a cook who memorized the menu but can't whip up anything new.

After creating a catalog of computer simulations of explosions with different matter, researchers ran a standard test: they shuffled the data and hid a portion — accuracy hit 70%. But when they excluded an entire class of matter, the program failed: predictions became worse than random guessing. The algorithm wasn't learning physics, just the specific contours of the waves.

When the task was simplified to determining the mass of the progenitor star, the program learned to generalize. This is a caution for future detectors: we need tests on never-before-seen models. Otherwise, a clever algorithm remains a trained cook, incapable of improvisation.