Retrieval head. Distinct from induction: rather than keying on local repetition to predict the next token, a retrieval head copies a semantically required span from far away to answer a query — defined by attention onto a known target, not by next-token completion. They are sparse and universal across models, and causally tied to factuality: knock them out and needle-in-a-haystack recall fails. Amber tier: the copy-from-distance mechanism is exact; the haystack is constructed.