Transferable and Local Evidence in Agent Reputation: Why a Missing Proof Must Not Travel

An agent that detects a compromised counterpart knows something other agents would benefit from knowing. Sharing it is reputation, and reputation systems are historically attackable because they aggregate assertions, which an adversary can author. A natural fix is to aggregate only facts an agent verified itself, by a check any other agent could reproduce. Continuity verifications, from the companion paper, are such facts. We argue that fix is not sufficient. When a counterpart presents a wrong proof, the observer holds a failing value any agent can recheck: a commission fault, with self-contained evidence. When a counterpart presents no proof, the observer has only its own view: an omission fault, for which Haeberlen and Kuznetsov proved no self-contained proof exists. The companion work measured legitimate counterparts producing exactly this signal in 60.7% of sessions under ordinary message reordering. A naive rule that propagates both network-wide turns the disqualification rule into a denial-of-service weapon. We measure the cost precisely: under the naive rule, one adversarial participant disqualifies an honest counterpart in 100.0% of trials, and ordinary 5% delivery loss disqualifies a legitimate counterpart with 20 correspondents in 63.0% of trials, with no adversary present. Under the revised rule, which propagates commission-class events and confines omission-class events to the observing agent, both failures fall to zero, structurally rather than statistically. We report the price of the correction in the same breath: an attacker who takes over an address and never emits a token is undetectable network-wide, measured at 0.0% versus 100.0% under the naive rule. We also measure what accumulation actually buys once a proof does propagate, between 58% and 90% advance protection for seed-holding peers, depending on contact frequency, in a stochastic network simulation, and state plainly what that simulation does not establish: that independently-operated agents would run the protocol faithfully. Related work by the author: Agent-Identity Continuity: Detecting Counterpart Takeover After Authentication Succeeds (DOI 10.5281/zenodo.22119416), which supplies the events aggregated here. And Context References Over Payloads: Authority Scoped to the Predicate, Not the Principal (DOI 10.5281/zenodo.21860668).