AgentValve caps Sybil-driven payment exposure by trust capacity

How can clients authorize payments when anyone can create cheap identities and coordinate feedback to make a provider look trustworthy?

The concern is that a reputation-based rule could let the amount at risk grow with the attacker's identities, rather than with trust capacity that honest participants actually extend.

AgentValve replaces global reputation with capacity-limited trust paths, atomic reservations, and compact on-chain verification to bound Sybil exposure.

The prototype was tested on five selected regions of an Ethereum-derived trust graph and on synthetic gateway graphs. Researchers compared AgentValve with two global rules, calibrating each rule so one identity authorized the same exposure as AgentValve's honest-to-adversarial cut capacity. They increased the number of added identities from 1 to 10,000.

AgentValve stayed at the fixed cut capacity C, while both global rules authorized mC in both settings. A constructed worst-case route also reached the formal H_max−1 depletion ratio, showing that the stated bounds can be tight under their intended conditions.

These observations support bounded exposure and bounded depletion, rather than immunity to every loss.

The cut result is a snapshot bound on unresolved and jointly admissible risk, not cumulative fraud loss; timeout resolution can permit repeated losses.

The depletion denominator counts gross protocol payments, so it does not establish an attacker's net economic cost after refunds, service value, or fees.