The strongest scientific answer is conditional. Context can induce a computation equivalent to a parameter update, and theoretical work maps this effect to low-rank feedforward changes. Yet exact, persistent conversion is architecture-dependent and unavailable in standard transformers. A convincing claim of stable learning therefore requires more than prompt-conditioned behaviour: researchers must show retention after context removal, controlled generalisation, resistance to interference, and an explicit write mechanism. The unresolved issue is whether these criteria define learning or merely a highly structured simulation of it.