Mechanistic Interpretability Take-Home — Sample-Aspect Double Descent

Reproduce sample-aspect double descent, then make the study fast enough to run The take-home this comes from was open-ended: design an experiment that shows double descent, explain the mechanism, propose a fix, present it. The workspace below is that study already half-built and quietly wrong — a small NumPy/SciPy repo you have to repair, finish and speed up. The setting is the simplest one that shows the phenomenon honestly: ridge regression on synthetic linear-Gaussian…