You already know: how to read and check a paper (Reading papers); ablations, seeds and significance (Research workflow); experiment tracking and reproducible pipelines (D11); training from scratch in NumPy and PyTorch (D4).
Why this matters for Atlas. Reproduction is how research engineers turn other people's claims into knowledge they can build on, and it's a standard interview exercise at research labs. It's harder than it looks: papers omit details that turn out to matter, defaults in libraries differ from the authors' choices, and a reproduction that "doesn't work" is usually a reproduction of something slightly different. This lesson gives a process: scope the claim, list and resolve ambiguities, build bottom-up with tests, reproduce the paper's setting before your own, debug mismatches systematically, and write a report people read. The running example reproduces double descent (Belkin et al., 2019) in NumPy, and shows how one library default makes it disappear.
1. Scope: which claim, at what cost
The question. What exactly are you reproducing?
Start from nothing. Pick one claim and the smallest experiment that tests it. "Reproduce the paper" is unbounded; "reproduce Figure 2: with random ReLU features and minimum-norm least squares, test error peaks when the number of features equals the number of training points, and falls again beyond it" is a plan. Decide up front what counts as success: the qualitative shape? The number within some tolerance? Start at the smallest scale where the effect should appear; scale up only after it does.