Who Stands Behind the Result? Vouching as the Missing Epistemic Value in Science with Agentic AI

Scientific results are not owned; they are vouched for. Every measurement that enters the record carries a chain of people who put their names behind it so that others can build on it. Agentic AI systems now write code, run simulations, and assemble analyses, inserting links into that chain to which no name attaches: output without an address. Hallucination research measures how often such output is false; vouching asks who answers for it. Writing from a constellation of cultural anthropology, artificial intelligence research, and particle physics, we argue that vouching, the circulating practice of standing behind what one passes on, is a load-bearing epistemic value that the AI-safety literature has neither named nor operationalized. Gift theory supplies the reason it cannot be delegated: a giver stands to be diminished or honoured by giving; an agent has nothing at stake. High-energy physics is where this epistemic value becomes observable through its rigorous documentation of approvals, so a check passing from a person to an agent inherits a record and can be dated. We identify four attachment points where humans enter agentic AI pipelines, asking, judging, signing, and supervising; propose three observables, substitution events, gate evidence, and revocability; and describe the small artefact, a substitution record, that would carry them. Our approach is design-anthropological: gates are to be built with the working groups and evaluated where approvals are already on record. A published vouching chain, we argue, is where human-machine collaboration in science becomes observable and governable. The paper closes on the constellation it was written from, offered as a repeatable form for a problem no single community can hold.