[!todo] Seed note. A starting point, not a finished note yet.
CACE stands for “Changing Anything Changes Everything,” named in Sculley et al.’s 2015 technical-debt paper. Because an ML model mixes all its signals into one set of learned weights, no input is independent: adding or removing a feature, changing a hyperparameter, or a shift in the input distribution moves the whole model, so a change that looks local can degrade unrelated predictions. This is why “improving” one feature can make the system worse, and why the usual software instinct to reason about a change in isolation fails for ML. The paper’s mitigations are containment (isolate models, combine via ensembles) and detection (monitor prediction differences) rather than cure. Seeded from Hidden Technical Debt in Machine Learning Systems.