We built a full instrument suite for reading the inside of trained neural networks — and it produced findings on the first day of operation. Everything is public, pre-registered, and reproducible. The setup, in one line: take any AI model's weights, transform them into a spectral basis (think: a prism for numbers), and compare against shuffled copies of the same numbers. Whatever signal survives can only come from where training placed the values — pure structure, not statistics. What we fou...