[!todo] Seed note. A starting point, not a finished note yet.
A correction cascade is a system-level ML trap named in Sculley et al.’s 2015 technical-debt paper. You have a model for problem A, then face a slightly different problem A’, and rather than solve A’ directly it’s faster to learn a small correction model on top of the first model’s output. That correction now depends on the original, so improving the original model can hurt overall accuracy. The cascade creates an improvement deadlock where making any single component better makes the system worse. Cascade several corrections and the coupling becomes intractable to reason about. The paper’s advice is to solve A’ directly by adding features to the original model, and to accept that a separate model is sometimes the cheaper long-run choice. Seeded from Hidden Technical Debt in Machine Learning Systems.