Retrieval Trust Gradient

Personal assistants earn their keep by retrieving the right context at the right moment, but retrieval quality is not uniform. A local assistant such as MyHandler can answer 'what did I quote Acme?' because it captured the screen, transcript, or inbox where that number appeared. Ask the same assistant about a client relationship that only exists in a hallway conversation and the recall confidence collapses. The Retrieval Trust Gradient ranks every task by how well the assistant's captured corpus actually covers it, then sets a delegation ceiling: high-confidence retrieval tasks go to the assistant, low-confidence ones stay with a human. For agencies, this matters because misjudging the gradient produces confident wrong answers in client-facing work. PHNTM One's on-device document Q&A with citations is a useful benchmark here, since citation-backed answers let a delivery lead verify before sending. Map the gradient per client retainer, not per tool.