Definition
Pretending some of an AI's notes were stolen, and seeing how much performance drops.
In Auto-Dreamer, a counterfactual subset-masking procedure that randomly removes synthesized memory entries and measures the resulting performance change.