Chain of Epistemic Custody: Nothing Crosses Between AI Agents Without Lineage

AI agents now pass claims to one another and act on them: they research, plan, buy, file and pay. This creates three gaps. An estimate produced by one agent can reach another as a fact, because nothing that travels with it records what it was. An agent can act on grounds that no one has checked. And when one model is asked to judge the work of models, it brings measured biases, including a preference for work like its own. This paper proposes the Chain of Epistemic Custody: nothing crosses between agents, neither a claim nor an action, without lineage. Its unit is the Information Passport, a record that travels with every material claim and states: its source; whether it is measured, estimated or inferred; its date and confidence; the unknowns its producer declares. The paper: sets six rules, including "no upgrade in transit" and "unknowns are inherited"; defines an independent arbiter that locks its criteria before seeing the work, decides by evidence where evidence exists, lets no single language model decide a decisive question, and never sits as a party or a source; defines fourteen reference criteria for judging, seven for the quality of information and seven for the quality of outputs, and judges agents only by the record of their judged outputs; lets the arbiter hear a council of experts that testifies independently first, then deliberates, and never votes on the verdict; states when information becomes knowledge: a claim reaches that standing only when an independent judgment finds five conditions met for a stated use (origin shown, standing honest, path shown, unknowns settled for the use, independently judged), and it loses that standing when one fails; an estimate can be known only as an estimate; has every verdict leave an auditable record, a visible fingerprint and, on request, a signed certificate (W3C Verifiable Credentials 2.0) with a printable verdict deed, attesting the conduct and result of the judgment, not the truth of the output; binds consequential actions resting on material unknowns to a recorded clearance at the highest conformance level; defines a judge-of-judges evaluation, with an independence rule for the standard's own editor, and a supplier rule for arbiters whose technology supplier is also a competing producer (disclosure, safeguards in every role, and measured preference for that supplier); positions the proposal against current work, and claims novelty only for the combination; sets out four falsifiable hypotheses and three pre-specified experiments: cascades, the power to act, and judging the judges. The proposal is specified as an open standard, the Epistemic Custody Standard (ECS-1). Mind Intelligence, Mind Engineering and Mind Governance are a framework introduced by Abdullah Eliyan (2026). Chain of Epistemic Custody and Information Passport are terms introduced by Abdullah Eliyan (2026) in GAIBRA Whitepaper No. 2 and developed in this paper.